Consideration of predicted small-molecule metabolites in computational toxicology
Consideration of predicted small-molecule metabolites in computational toxicology
复制标题
DOI：
10.1039/d1dd00018g
复制
发表时间：
2022-04-11
期刊：
影响因子：
--
通讯作者：
Kirchmair, Johannes
中科院分区：
文献类型：
--
作者：
de Lomana, Marina Garcia;Svensson, Fredrik;Kirchmair, Johannes
外源代谢已经发展成为生物体对抗潜在有害化学物质或化合物的关键保护系统，通常不存在于特定生物体中。该系统的主要目的是将外源药物化学转化为可通过肾脏或胆道排出的代谢物。然而，在少数情况下，形成的代谢物是有毒的，有时甚至比母体化合物更有毒。因此，明确地考虑外源代谢对了解化合物的毒性具有重要意义。然而，大多数现有的毒性预测计算方法没有明确考虑代谢，目前尚不清楚考虑（预测的）代谢物能在多大程度上改善毒性预测。为了研究预测代谢如何有助于增强毒性预测，我们探索了许多不同的策略，以整合来自最先进的代谢物结构预测器的预测和用于毒性预测的现代机器学习方法。我们在5个毒理学终点和试验中测试了集成模型，包括体外和体内遗传毒性试验（AMES和MNT），两个器官毒性终点（DILI和DICC）和皮肤致敏试验（LLNA）。总体而言，纳入代谢数据对模型性能的改善很小（F1得分最高为+0.04，mcs得分最高为+0.06）。一般来说，通过对母体化合物预测的毒性概率和对任何代谢物预测的最大毒性概率进行平均，可以获得最佳性能。此外，将代谢物结构作为模型训练的进一步输入分子，略微改善了通过这种平均方法获得的毒性预测。然而，代谢系统的高度复杂性和可能代谢物的不确定性显然限制了在毒性预测中考虑预测代谢物的益处。探索在毒性预测的机器学习模型中包含代谢信息的计算方法。
Xenobiotic metabolism has evolved as a key protective system of organisms against potentially harmful chemicals or compounds typically not present in a particular organism. The system's primary purpose is to chemically transform xenobiotics into metabolites that can be excreted via renal or biliary routes. However, in a minority of cases, the metabolites formed are toxic, sometimes even more toxic than the parent compound. Therefore, the consideration of xenobiotic metabolism clearly is of importance to the understanding of the toxicity of a compound. Nevertheless, most of the existing computational approaches for toxicity prediction do not explicitly take metabolism into account and it is currently not known to what extent the consideration of (predicted) metabolites could lead to an improvement of toxicity prediction. In order to study how predictive metabolism could help to enhance toxicity prediction, we explored a number of different strategies to integrate predictions from a state-of-the-art metabolite structure predictor and from modern machine learning approaches for toxicity prediction. We tested the integrated models on five toxicological endpoints and assays, including in vitro and in vivo genotoxicity assays (AMES and MNT), two organ toxicity endpoints (DILI and DICC) and a skin sensitization assay (LLNA). Overall, the improvements in model performance achieved by including metabolism data were minor (up to +0.04 in the F1 scores and up to +0.06 in MCCs). In general, the best performance was obtained by averaging the probability of toxicity predicted for the parent compound and the maximum probability of toxicity predicted for any metabolite. Moreover, including metabolite structures as further input molecules for model training slightly improved the toxicity predictions obtained by this averaging approach. However, the high complexity of the metabolic system and associated uncertainty about the likely metabolites apparently limits the benefit of considering predicted metabolites in toxicity prediction.Exploration of computational approaches for including metabolism information in machine learning models for toxicity prediction.