An adverse outcome pathway (AOP) is a description of the sequence of causally linked events, spanning multiple levels of biological organization, required to produce a toxic effect when an organism is exposed to a stressor such as a chemical. In essence, an AOP is a depiction of a complex toxicological process in a simplified, stepwise, sequential format. An AOP tries to describe the responses of complex biological systems whose behavior is intrinsically difficult to model because of the many relationships and interactions between the various components of the system and because of the emergent properties that often arise as a result of such interactions. An important question, therefore, is whether the AOP framework itself should, by design, also be a complex system to be able to capture the toxicological reality or whether a simplified representation of the perturbed biology is indeed sufficient. This question has been raised ever since the introduction of the AOP framework in 2010 (Ankley et al., 2010). Some argue that the “linear” nature of AOPs is too simple in terms of conceptual model design and data type availability to adequately capture the complexity of any realistic toxicological scenario, even when considering AOP networks that are formed by connecting different linear AOPs. Others feel that the AOP framework is too complex and too overwhelming to be useful. Development of an AOP can be perceived as a daunting task, requiring the knowledge and evidence to connect all the dots between the different events, including establishing causality and essentiality, documenting dose and time concordance, and so on. Thus, AOPs are often perceived as being too simple and too complex at the same time, raising concerns that they may in fact slow down hazard and risk assessment instead of being a catalyst for supporting 21st-century toxicology.

Of course, this paradox is not new, and it is certainly not unique to the AOP framework. Humanity has been studying complex systems for centuries. Intuitively one tries to reduce complexity to be able to understand reality and deal with it in a practical manner. At the same time there is a natural tendency to add ideas and functionality, and therefore complexity, to almost anything we design to increase its relevance and utility. Researchers in complex systems view the main task of modeling to capture, rather than reduce, the complexity of the systems of interest; but the question remains what the level of complexity of the model should be to achieve this. To further explore the paradox of “complex simplicity,” it is helpful to consider the relationship between complexity and simplicity and to distinguish between different types of simplicity. While simplicity and complexity are usually perceived as extreme ends of the same spectrum, they can also be seen as interchangeable. True simplicity can indeed be utterly complex. Steve Jobs, the founder and late chief executive officer of Apple, said on this matter,

In other words, the development of a new idea or paradigm evolves through different phases, starting from a pragmatically simple idea, leading to an intermediate phase of necessary added complexity, and ending in a final state of what is sometimes referred to as “informed simplicity,” gradually increasing its utility throughout the process (Figure 1). Of course, Jobs made his statement in the context of Apple products, the development of the iPhone—a highly complex but very simple-to-use product—out of the analog telephone via “smart” intermediate technology such as BlackBerry-type devices being a good example. The idea of aiming for evolved, informed simplicity applies equally well to the history and future of the AOP framework.

At their inception AOPs were intended to be pragmatically simple. Individual linear AOPs, in particular, were conceived as a deliberate simplification of complex biology to serve as a functional unit of development intended to support regulatory decision-making. For example, it was explicitly prescribed that AOPs should contain only a limited number of essential, measurable events leading to the relevant toxicity endpoint and should not necessarily provide a comprehensive description of every aspect of the biology involved. And even though AOP networks were envisioned to be able to better capture the complexity of biological systems, they were initially seen as relatively simple constructs as well, only made up of the information contained within the constituent AOPs with no additional layers of network-specific properties, data, or metadata.

It soon became apparent that there was a need to expand the AOP development toolbox. Over the last decade and through the present, the AOP framework has been gradually refined to suit the evolving needs of AOP developers and users, and a variety of features and properties have been added to accommodate a number of perceived challenges and shortcomings. For example, it was considered how feedback loops and modulating factors could be integrated into the framework. A lot of thought and consideration went into the best way to define and describe an AOP's domain of applicability, for example, in terms of taxonomic scope (LaLone et al., 2013), life stage specificity, or sex specificity. A system of layers was proposed to add information to AOPs and AOP networks without directly affecting the underlying pragmatic simplicity of the construct. Filters were then proposed as a way to reduce the ever-growing complexity of large AOP networks (Knapen et al., 2018). The use of ontology terms was introduced to standardize the description of different pathway components (Ives et al., 2017), systematic approaches were explored for collecting the available scientific evidence supporting an AOP, and Bradford Hill–based criteria were established for evaluating that evidence (Becker et al., 2015). Methods were developed to analyze, extract, and benchmark the information that is contained within complex AOP networks (Pollesch et al., 2019; Villeneuve et al., 2018). While many of these framework improvements are now well established, not all are yet fully developed or understood. Developing the AOP framework does remain an active area of science, and the framework continues to evolve. The progress that has been made in the past few years is significant however, and it is likely that we are almost in the middle of the simplicity continuum (see Figure 1). What we need now is a way to bring all the available information, technology, and concepts together. Most importantly, there is a need for a new, advanced data model to link and integrate all the data and associated metadata that make up AOPs and AOP networks and a new conceptual way for a user to interact with these data components to explore and interrogate the AOP ecosystem. Fortunately, the necessary steps in this direction are currently being taken: a new data model is indeed under development, and the AOP-wiki is evolving into a mature and sophisticated interface for both AOP developers and users.

This brings us back to our question of whether AOPs themselves, or the AOP framework itself, should be a complex system by design. The answer is that the AOP framework needs to achieve a state of informed simplicity where it still looks, feels, and handles as simply and elegantly as originally designed but at the same time incorporates more of the great complexities that make up life. Many of the ingredients are in place, much like most of the individual pieces of technology required to build the iPhone existed just before it was invented. The only thing that is left to do is to search for the key, the underlying, principle of the problem—and come up with an elegant, really beautiful solution that works.

不良结果通路 (AOP) 是对因果关联事件序列的描述，跨越多个生物组织水平，当有机体暴露于化学物质等压力源时产生毒性作用所需。本质上，AOP 是以简化的、逐步的、连续的格式描述复杂的毒理学过程。AOP 试图描述复杂生物系统的响应，这些系统的行为本质上难以建模，因为系统的各个组件之间存在许多关系和相互作用，并且由于这种相互作用经常会出现紧急特性。因此，一个重要的问题是 AOP 框架本身是否应该按照设计，也是一个能够捕捉毒理学现实的复杂系统，或者扰动生物学的简化表示是否确实足够。自 2010 年引入 AOP 框架以来，这个问题就一直被提出（Ankley et al., 2010）。一些人认为，AOP 的“线性”性质在概念模型设计和数据类型可用性方面过于简单，无法充分捕捉任何现实毒理学场景的复杂性，即使考虑通过连接不同线性 AOP 形成的 AOP 网络也是如此。其他人则认为 AOP 框架过于复杂，过于庞大而无用。AOP 的开发可以被视为一项艰巨的任务，需要知识和证据来连接不同事件之间的所有点，包括建立因果关系和必要性，记录剂量和时间的一致性等等。因此，AOP 通常被认为既过于简单又过于复杂，引发了人们的担忧，即它们实际上可能会减缓危害和风险评估，而不是成为支持 21 世纪毒理学的催化剂。

当然，这个悖论并不新鲜，当然也不是 AOP 框架独有的。几个世纪以来，人类一直在研究复杂的系统。直觉上，人们试图降低复杂性，以便能够理解现实并以实际的方式处理它。同时，我们有一种自然的趋势，即为我们设计的几乎任何东西添加想法和功能，从而增加复杂性，以增加其相关性和实用性。复杂系统的研究人员认为建模的主要任务是捕捉而不是降低感兴趣系统的复杂性；但问题仍然是模型的复杂程度应该是多少才能实现这一目标。为了进一步探索“复杂简单”的悖论，考虑复杂性和简单性之间的关系以及区分不同类型的简单性是有帮助的。虽然简单性和复杂性通常被视为同一范围的极端，但它们也可以被视为可互换的。真正的简单确实可以非常复杂。苹果创始人、已故首席执行官史蒂夫乔布斯就此事表示，

换句话说，一个新的想法或范式的发展经历了不同的阶段，从一个实用的简单想法开始，到一个必要的增加复杂性的中间阶段，最后到一个有时被称为“知情简单”的最终状态。 ，”在整个过程中逐渐增加其效用（图 1）。当然，乔布斯在苹果产品的背景下发表了他的声明，iPhone 的开发——一种高度复杂但非常易于使用的产品——通过“智能”中间技术（如 BlackBerry 类型的设备）摆脱了模拟电话的影响。一个很好的例子。以进化的、知情的简单为目标的想法同样适用于 AOP 框架的历史和未来。

AOP 一开始就旨在简单实用。特别是单个线性 AOP，被认为是对复杂生物学的有意简化，以作为旨在支持监管决策的开发功能单元。例如，明确规定 AOP 应仅包含有限数量的导致相关毒性终点的必要的、可测量的事件，并且不一定提供所涉及生物学的各个方面的全面描述。尽管 AOP 网络被设想为能够更好地捕捉生物系统的复杂性，但它们最初也被视为相对简单的构造，仅由组成 AOP 中包含的信息组成，没有额外的网络特定属性层、数据或元数据。

很快就发现需要扩展 AOP 开发工具箱。在过去十年和现在，AOP 框架已经逐渐完善，以适应 AOP 开发人员和用户不断变化的需求，并且添加了各种特性和属性以适应许多感知到的挑战和缺点。例如，考虑了如何将反馈回路和调节因素整合到框架中。在定义和描述 AOP 的适用领域的最佳方式中，我们进行了很多思考和考虑，例如，在分类范围方面（LaLone et al.,  2013)、生命阶段特异性或性别特异性。提出了一种分层系统来向 AOP 和 AOP 网络添加信息，而不会直接影响构造的底层实用简单性。然后提出了过滤器作为降低大型 AOP 网络日益增长的复杂性的一种方法（Knapen 等人，  2018 年）。引入了本体术语的使用来标准化不同途径组件的描述（Ives 等人，  2017 年），探索了收集支持 AOP 的可用科学证据的系统方法，并建立了基于 Bradford Hill 的标准来评估该证据（贝克尔等人，  2015）。开发了一些方法来分析、提取和基准测试复杂 AOP 网络中包含的信息（Pollesch 等人，  2019 年；Villeneuve 等人，  2018 年））。虽然这些框架改进中的许多现在已经很好地建立起来，但并不是所有的都得到了充分的开发或理解。开发 AOP 框架确实仍然是一个活跃的科学领域，并且该框架还在不断发展。然而，过去几年取得的进展是显着的，我们很可能几乎处于简单连续体的中间（见图 1）。我们现在需要的是一种将所有可用信息、技术和概念结合在一起的方法。最重要的是，需要一种新的、高级的数据模型来链接和集成构成 AOP 和 AOP 网络的所有数据和相关元数据，并需要一种新的概念方式供用户与这些数据组件交互以探索和询问AOP 生态系统。幸运的是，目前正在朝着这个方向采取必要的步骤：

这让我们回到了 AOP 本身或 AOP 框架本身是否应该是设计上的复杂系统的问题。答案是 AOP 框架需要达到一种明智的简单状态，它的外观、感觉和处理仍然像最初设计的那样简单和优雅，但同时包含更多构成生活的巨大复杂性。许多要素都已准备就绪，就像制造 iPhone 所需的大多数单独技术在 iPhone 发明之前就已经存在一样。剩下要做的唯一一件事就是寻找问题的关键、根本、原理——并提出一个优雅的、非常漂亮的有效解决方案。