Conclusions via unique predictions obtained despite unidentifiability - new definitions and a general method
Conclusions via unique predictions obtained despite unidentifiability - new definitions and a general method
复制标题
DOI：
10.1111/j.1742-4658.2012.08725.x
复制
发表时间：
2012-09-01
期刊：
影响因子：
5.4
通讯作者：
Cedersund, Gunnar
中科院分区：
文献类型：
--
作者：
Cedersund, Gunnar
人们经常预测，基于模型的数据分析将彻底改变生物学，就像它对物理学和工程学一样。在这种分析中广泛使用的工具是假设检验，其重点是模型拒绝。然而，一个系统生物学模型是不被拒绝的事实往往是一个相对较弱的声明，因为这样的模型通常是高度过度参数化相对于可用的数据，参数和预测可能因此是任意不确定的。为此，我们正式定义和分析的核心预测的概念。核心预测是一个唯一识别的属性，如果给定的模型结构是为了解释数据，即使单个参数是非唯一识别的，也必须满足。它表明，这样的预测是一个强有力的结论作为拒绝。此外，介绍了一种新的岩心预测分析方法，该方法有利于特定模型属性的不确定性，因为该方法仅表征相关方向上可接受参数的空间。这避免了与先前提出的方法所使用的通用表征相关联的维数灾难。实例分析表明，新方法在实际可辨识性方面与轮廓似然法相当，从而将轮廓似然法推广到更一般的可观测性问题。如果使用，本文提出的概念和方法可以区分结论和仅仅是建议，这有望有助于系统生物学分析中更合理的信心。
It is often predicted that model-based data analysis will revolutionize biology, just as it has physics and engineering. A widely used tool within such analysis is hypothesis testing, which focuses on model rejections. However, the fact that a systems biology model is non-rejected is often a relatively weak statement, as such models usually are highly over-parametrized with respect to the available data, and both parameters and predictions may therefore be arbitrarily uncertain. For this reason, we formally define and analyse the concept of a core prediction. A core prediction is a uniquely identified property that must be fulfilled if the given model structure is to explain the data, even if the individual parameters are non-uniquely identified. It is shown that such a prediction is as strong a conclusion as a rejection. Furthermore, a new method for core prediction analysis is introduced, which is beneficial for the uncertainty of specific model properties, as the method only characterizes the space of acceptable parameters in the relevant directions. This avoids the curse of dimensionality associated with the generic characterizations used by previously proposed methods. Analysis on examples shows that the new method is comparable to profile likelihood with regard to practical identifiability, and thus generalizes profile likelihood to the more general problem of observability. If used, the concepts and methods presented herein make it possible to distinguish between a conclusion and a mere suggestion, which hopefully will contribute to a more justified confidence in systems biology analyses.