Imprecision and importance: Probabilistic graphical models in toxicology

为了评估一种新化学品的安全性，现有的毒理学协议通常需要至少4年的研究，300万至400万英镑的资金，以及对4000多只动物的实验。这是低效的，道德上有问题的，并且未能利用最近的生物，数学和计算进步。我们计划使用先进的计算和统计方法来研究如何改进评估过程。更明确地说，我们将使用贝叶斯网络来利用现有数据，以确定有效毒理学评估所需的关键研究，并量化应该容忍的不精确程度。贝叶斯方法使我们能够利用现有的关于毒性的知识，即使这些知识是不完整或不准确的，并使用它来更好地预测未来的化学品。由于不确定性和不精确性可以很自然地建立在模型中（事实上，它们是必要的），因此更容易对毒理学风险做出数据驱动的概率性陈述，其输出将提供对现有协议中每个元素的价值的严格量化。在此过程中，我们将试图回答这样一个问题，即来自50多种经过充分研究的有毒物质的详细数据在多大程度上可以证明在未来化学品所需的毒理学测试中应用3R（动物实验的减少、改进和替代）是合理的。明确地说，如果我们的模型表明可以从较小的测试组合中获得足够的精度，则将减少用于测试的动物数量;如果我们的模型通过对广谱数据的整体同化来提高效率，则动物测试协议的改进将成为可能;在适当的情况下，在给定化学品测试策略的早期，可以合理和量化地证明动物测试的合理性。

To evaluate the safety of a new chemical existing toxicological protocols typically require at least 4 years of research, £3-4 million in funding, and experiments on over 4000 animals. This is inefficient, ethically questionable, and fails to exploit recent biological, mathematical and computational advances. We plan to use advanced computational and statistical methods to investigate how the evaluation process can be improved.More explicitly, we will use Bayesian networks to exploit existing data both to identify the key studies necessary for efficient toxicological assessment, and to quantify what levels of imprecision should be tolerated. Bayesian methods allow us to take existing knowledge about toxicity, even if this knowledge is incomplete or inaccurate, and use it to make better predictions for future chemicals. Because uncertainty and imprecision can be built very naturally into the models (indeed, they are necessary), it becomes easier to make data-driven probabilistic statements about toxicological risk.The outputs will provide a rigorous quantification of the value of each element of existing protocols. In doing so, we will seek to answer the question of how far the detailed data from a panel of 50+ well-studied toxicants can justify applying the 3Rs (the reduction, refinement and replacement of animal experiments) in the toxicological testing necessary for future chemicals. Explicitly, reduction of the number of animals used for testing will be achieved where our models indicate that enough precision can be derived from a smaller battery of tests; refinement of animal testing protocols will become possible where our models idenitify efficiencies through the holistic assimilation of broad-spectrum data; and where appropriate replacement of animal tests by can be rationally and quantifiably justified early in a given chemical's testing strategy.