This article is part of the Artificial Intelligence in Drug Discovery special issue. We are pleased to introduce the Special Issue “Artificial Intelligence in Drug Discovery” highlighting the emerging role of artificial intelligence (AI) in pharmaceutical research. A focal point of this issue is illuminating how AI approaches are beginning to impact the practice of drug discovery. The Special Issue contains articles and perspectives that view AI in drug discovery from different angles. First, we thank our authors for their high-quality and thematically diverse contributions. In addition, we thank many reviewers of submissions to this Special Issue who often have carefully evaluated manuscripts at short notice. Finally, we gratefully acknowledge the editorial office staff for their continuous support. Of note, a significant number of the many submissions to this Special Issue could not be further considered because they did not meet the Author Guidelines and general acceptance criteria for publication of computational studies in the Journal of Medicinal Chemistry. The papers in this issue cover a variety of method development efforts and practical applications that provide us with a flavor of how AI is entering the drug discovery arena. Of course, machine learning methods have already been applied for more than two decades in cheminformatics and computational medicinal chemistry, but deep learning has more recently become a hot topic in many areas of science including chemistry. The Special Issue pays tribute to these developments. In addition to providing a number of applications of AI in drug discovery, the Special Issue contains papers that address several key features in the field, as highlighted below. Molecular Representation. A machine learning method creates a model that is used to establish relationships between input data and an observable end point. In medicinal chemistry, we typically model the relationship between chemical structure and physical properties or biological activity. A key component in this process is the representation that is used to map a molecular structure into a form that can be processed by a machine learning algorithm. Until recently, molecules were often represented as vectors encoding the presence or absence of substructures of structural patterns in a molecule. In recent years, a number of groups have developed methods that use deep learning to create new “learned representations”. While the predictive power of these learned representations is still an open question, they have shown promising initial results. One of the perspectives in this issue provides an excellent introduction to the motivations behind and applications of learned representations (Chuang et al.; DOI: 10.1021/acs.jmedchem.0c00385). In addition, this issue features two articles on the topic of learned representations. One paper describes the application of a technique known as graph attention networks, which enable a neural network to focus on the most important features (Xiong et al.; DOI: 10.1021/acs.jmedchem.9b00959). The other paper reports the benchmarking of a novel molecular representation on a large set of pharmaceutical ADME data (Feinberg et al.; DOI: 10.1021/acs.jmedchem.9b02187). Model Interpretation. One drawback to many machine learning approaches is that they mostly are “black box predictors”. Accordingly, in a drug discovery setting, one inputs a chemical structure and receives a result without any explanation of how or why the prediction was generated. Ideally, we would like to have machine learning models that could be interpreted by human users. The explanations produced by these models would serve two purposes. First, a user could examine the explanation, confirm that it agrees with theoretical and experimental foundations, and establish some degree of confidence in the prediction. Second, the explanation of the model could provide clues into the mechanistic drivers behind the biological activity being modeled and provide inspiration for the design of new molecules. The Special Issue contains a paper describing new methods that enable machine learning models to provide interpretable results for chemical data sets, regardless of the algorithms that are used (Rodríguez-Pérez et al.; DOI: 10.1021/acs.jmedchem.9b01101). Recommendation Systems. Computer systems that provide recommendations have become part of our everyday lives. For example, e-commerce sites provide recommendations based on our purchasing history. Online streaming sites recommend music and videos that we may enjoy. A paper in this issue extends this concept to the medicinal chemistry laboratory. The authors describe how the concept of “people who bought this also bought this” can be extended to recommending routes for organic synthesis, three-dimensional structures of similar compounds, and assays that may provide additional insights (Rohall et al.; DOI: 10.1021/acs.jmedchem.9b02130). Reaction Design. One of the areas in chemistry where AI is currently making headway is predicting and modeling new chemical reactions and synthetic routes. A perspective in this issue highlights recent developments in this emerging area of research and provides an outlook (Struble et al.; DOI: 10.1021/acs.jmedchem.9b02120). Generative Models. Despite more than 30 years of advances in computational chemistry, many, if not most, of the ideas for new molecules in a drug discovery program originate from the imagination and ingenuity of a medicinal chemist. Beginning in the 1990s, a number of groups produced computer programs for performing de novo molecular design. These programs often (but not always) operated by “growing” an existing molecule in the context of a protein binding site. However, while there were some stories of success from de novo design, the technique failed to achieve mainstream adoption. Over the past few years, we have seen the rise of a related technique known as generative modeling. This field, which has its origins in language models and image generation, takes as input a set of molecular structures, which are encoded as a continuous low-dimensional representation. This representation can then be decoded to generate new, often novel, molecules. However, chemists may question the ability of such a system to learn the actual chemistry necessary to generate drug-like molecules. One of the papers in this issue provides an investigation of this potential caveat by evaluating the scope of the actual chemistry learned by a generative model (Grebner et al.; DOI: 10.1021/acs.jmedchem.9b02044). Perspectives. The Special Issue includes a number of important perspectives on the role of AI in drug discovery. One of these papers explores the broader theme of interactions between chemists and AI in depth (Griffen et al.; DOI: 10.1021/acs.jmedchem.0c00163), while another focuses on the impact of AI in synthesis (Struble et al.; DOI: 10.1021/acs.jmedchem.9b02120; vide supra). Many applications of AI in our field are still constrained by the limited availability of data, which is also addressed in a perspective discussing transfer learning methods that can be used to leverage knowledge gained in related projects to accelerate new efforts. Practical Impact. We also note that only very few of the papers appearing in this Special Issue showcase real-world applications of AI that currently impact drug discovery. We are certainly encouraged by in-house evaluation of these methods by those on the front lines in pharma, as highlighted in one of the contributions (Rohall et al.; DOI: 10.1021/acs.jmedchem.9b02130; vide supra), and we emphasize the need for these methods to be put to the test in more “risky” scenarios. In particular, this means that AI applications, and especially predictive models, need to have “skin in the game” and directly influence the selection and prioritization of compounds in high-profile drug discovery programs. However, demonstrating the impact of AI at that level is still a rare event. Importantly, as long as “prospective” evaluations only consider the choices made by chemists (rather than the choices made by the algorithms), the real impact of these methods will be difficult to assess and, importantly, nearly impossible to improve. From this point of view, the field is still wide open to advance AI in drug discovery beyond the conceptual level and demonstrate the ability of smart algorithms to consistently design novel chemical matter going beyond a chemist’s imagination. To these ends, the contributions contained in this Special Issue are also considered as an encouragement to move forward “into the great wide open” (Tom Petty & the Heartbreakers, 1991). The papers certainly provide a realistic impression of the current state-of-the-art. It is sincerely hoped that the readers of the Journal of Medicinal Chemistry will enjoy this Special Issue covering a hot topic in science from a drug discovery perspective. Views expressed in this editorial are those of the authors and not necessarily the views of the ACS. This article has not yet been cited by other publications.

中文翻译：

---

药物发现中的人工智能：大开放。

本文是 药物发现中的人工智能特刊。我们很高兴介绍“药物发现中的人工智能”特刊，着重介绍人工智能（AI）在药物研究中的新兴作用。这个问题的重点是阐明AI方法如何开始影响药物发现的实践。特刊包含从不同角度看待药物研发中的AI的文章和观点。首先，我们感谢作者的高质量和主题多样化的贡献。此外，我们还要感谢许多对此特刊投稿的审稿人，他们经常在短时间内仔细评估稿件。最后，我们非常感谢编辑部工作人员的不断支持。值得注意的是药物化学杂志。本期的论文涵盖了各种方法开发工作和实际应用，为我们提供了AI如何进入药物发现领域的味道。当然，机器学习方法已经在化学信息学和计算药物化学中应用了二十多年，但是深度学习最近已成为包括化学在内的许多科学领域的热门话题。特刊对这些事态发展表示敬意。除提供AI在药物发现中的许多应用外，特刊还包含针对该领域几个关键特征的论文，如下所述。分子表示。机器学习方法创建一个模型，该模型用于建立输入数据和可观察端点之间的关系。在药物化学中，我们通常对化学结构与物理性质或生物活性之间的关系进行建模。此过程中的关键组成部分是表示形式，用于将分子结构映射为可以由机器学习算法处理的形式。直到最近，分子通常被表示为编码分子中结构模式的亚结构存在或不存在的载体。近年来，许多小组开发了使用深度学习来创建新的“学习的表示形式”的方法。这些学习的表示的预测能力仍然是一个悬而未决的问题，但它们已经显示出令人鼓舞的初步结果。本期中的一种观点很好地介绍了学习表示的背后动机和应用（Chuang等； DOI：10.1021 / acs.jmedchem.0c00385）。此外，本期还将刊登两篇有关学习表示的文章。一篇论文描述了一种称为图注意力网络的技术的应用，该技术使神经网络能够专注于最重要的功能（Xiong等人； DOI：10.1021 / acs.jmedchem.9b00959）。另一篇论文报道了在大量药物ADME数据上对新型分子表示的基准测试（Feinberg等； DOI：10.1021 / acs.jmedchem.9b02187）。本期有两篇关于学习的表示形式的文章。一篇论文描述了一种称为图注意力网络的技术的应用，该技术使神经网络能够专注于最重要的功能（Xiong等人； DOI：10.1021 / acs.jmedchem.9b00959）。另一篇论文报道了在大量药物ADME数据上对新型分子表示的基准测试（Feinberg等； DOI：10.1021 / acs.jmedchem.9b02187）。本期有两篇关于学习的表示形式的文章。一篇论文描述了一种称为图注意力网络的技术的应用，该技术使神经网络能够专注于最重要的功能（Xiong等人； DOI：10.1021 / acs.jmedchem.9b00959）。另一篇论文报道了在大量药物ADME数据上对新型分子表示的基准测试（Feinberg等； DOI：10.1021 / acs.jmedchem.9b02187）。模型解释。许多机器学习方法的缺点之一是它们大多是“黑匣子预测器”。因此，在药物发现环境中，人们输入化学结构并接收结果，而无需任何解释如何或为什么生成预测。理想情况下，我们希望拥有可以由人类用户解释的机器学习模型。这些模型产生的解释有两个目的。首先，用户可以检查解释，确认它与理论和实验基础相符，并对预测建立一定的信心。其次，对该模型的解释可以为正在被建模的生物活动背后的机理驱动器提供线索，并为新分子的设计提供灵感。推荐系统。提供建议的计算机系统已成为我们日常生活的一部分。例如，电子商务网站会根据我们的购买历史记录提供建议。在线流媒体网站会推荐我们可能喜欢的音乐和视频。本期的一篇论文将此概念扩展到了药物化学实验室。作者描述了“购买此物的人也购买了此物”的概念如何可以扩展到推荐有机合成路线，类似化合物的三维结构以及可能提供更多见解的测定方法（Rohall等； DOI：10.1021）。 /acs.jmedchem.9b02130）。反应设计。AI目前正在取得进展的化学领域之一是预测和建模新的化学反应和合成路线。本期杂志的观点突出了这一新兴研究领域的最新发展，并提供了展望（Struble等； DOI：10.1021 / acs.jmedchem.9b02120）。生成模型。尽管计算化学取得了30多年的进步，但在药物发现计划中，许多（即使不是大多数）新分子的构想也源于药物化学家的想象力和独创性。从1990年代开始，许多小组生产了用于进行从头分子设计的计算机程序。这些程序通常（但不总是）通过在蛋白质结合位点的背景下“生长”现有分子来运行。但是，尽管有一些从头设计成功的故事，但该技术未能获得主流采用。在过去的几年中，我们看到了一种称为生成建模的相关技术的兴起。该领域起源于语言模型和图像生成，将一组分子结构作为输入作为输入，这些分子结构被编码为连续的低维表示。然后可以对该表示进行解码以生成新的，通常是新颖的分子。但是，化学家可能会质疑这种系统学习生成药物样分子所必需的实际化学反应的能力。本期的一篇论文通过评估由生成模型学到的实际化学作用的范围，对这一潜在的警告进行了研究（Grebner等人； DOI：10.1021 / acs.jmedchem.9b02044）。观点。特刊包括有关AI在药物发现中的作用的许多重要观点。其中一篇论文深入探讨了化学家与AI之间相互作用的更广泛主题（Griffen等人； DOI：10.1021 / acs.jmedchem.0c00163），而另一篇论文则着重探讨了AI在合成中的影响（Struble等人； DOI ：10.1021 / acs.jmedchem.9b02120;参见上文）。人工智能在我们领域中的许多应用仍然受到数据有限的限制，这在讨论转移学习方法的观点中得到了解决，转移学习方法可用于利用在相关项目中获得的知识来加速新的工作。实际影响。我们还注意到，本期特刊中只有极少数的论文展示了当前影响药物发现的AI的实际应用。正如其中一项贡献中所强调的那样（Rohall等人； DOI：10.1021 / acs.jmedchem.9b02130；见上文），我们当然对药房前线人员对这些方法的内部评估感到鼓舞。强调需要在更“危险”的情况下对这些方法进行测试。特别是，这意味着AI应用程序，尤其是预测模型，必须具有“游戏中的皮肤”的特性，并直接影响备受瞩目的药物发现程序中化合物的选择和优先级。但是，在此级别上展示AI的影响仍然是罕见的事件。重要的，只要“前瞻性”评估仅考虑化学家做出的选择（而不是算法做出的选择），这些方法的实际影响将难以评估，而且重要的是几乎无法改善。从这个角度来看，该领域仍然是广阔的空间，可以使AI在药物发现方面超越概念水平，并证明智能算法能够始终如一地设计出化学家无法想象的新颖化学物质。为此，本期特刊所载的贡献也被视为鼓励“迈向大开”（Tom Petty＆Heartbreakers，1991）。这些论文无疑为当前的最新技术提供了现实的印象。衷心希望广大读者 这些方法的实际影响将很难评估，而且重要的是几乎无法改善。从这个角度来看，该领域仍然是广阔的空间，可以使AI在药物发现方面超越概念水平，并证明智能算法能够始终如一地设计出化学家无法想象的新颖化学物质。为此，本期特刊所载的贡献也被视为鼓励“迈向大开”（Tom Petty＆Heartbreakers，1991）。这些论文无疑为当前的最新技术提供了现实的印象。衷心希望广大读者 这些方法的实际影响将很难评估，而且重要的是几乎无法改善。从这个角度来看，该领域仍然是广阔的空间，可以使AI在药物发现方面超越概念水平，并证明智能算法能够始终如一地设计出化学家无法想象的新颖化学物质。为此，本期特刊所载的贡献也被视为鼓励“迈向大开”（Tom Petty＆Heartbreakers，1991）。这些论文无疑为当前的最新技术提供了现实的印象。衷心希望广大读者 该领域仍然是广阔的空间，可以使AI在药物发现方面超越概念水平，并展示智能算法始终如一地设计出化学家无法想象的新颖化学物质的能力。为此，本期特刊所载的贡献也被视为鼓励“迈向大开”（Tom Petty＆Heartbreakers，1991）。这些论文无疑为当前的最新技术提供了现实的印象。衷心希望广大读者 该领域仍然是广阔的空间，可以使AI在药物发现方面超越概念水平，并展示智能算法始终如一地设计出化学家无法想象的新颖化学物质的能力。为此，本期特刊所载的贡献也被视为鼓励“迈向大开”（Tom Petty＆Heartbreakers，1991）。这些论文无疑为当前的最新技术提供了真实的印象。衷心希望广大读者 本期特刊所载的贡献也被视为鼓励“迈向大开放”（Tom Petty＆Heartbreakers，1991）。这些论文无疑为当前的最新技术提供了现实的印象。衷心希望广大读者 本期特刊所载的贡献也被视为鼓励“迈向大开放”（Tom Petty＆Heartbreakers，1991）。这些论文无疑为当前的最新技术提供了真实的印象。衷心希望广大读者《药物化学杂志》将享受本期特刊，从药物发现的角度探讨科学领域的热点话题。本社论中表达的观点只是作者的观点，不一定是ACS的观点。本文尚未被其他出版物引用。