AI for Organic and Polymer Synthesis

英文原文与中文译文逐段对照 · English original with Chinese translation

Xin Hong ${}^{1 * }$ , Qi Yang ${}^{2 * }$ , Kuangbiao Liao ${}^{3 * }$ , Jianfeng Pei ${}^{4 * }$ , Mao Chen ${}^{5 * }$ , Fanyang Mo ${}^{6,8 * }$ , Hua Lu ${}^{7 * }$ , Wen-Bin Zhang ${}^{7,8 * }$ , Haisen Zhou ${}^{7}$ , Jiaxiao Chen ${}^{4 * }$ , Lebin Su ${}^{3}$ , Shuo-Qing Zhang ${}^{1}$ , ${\text{ Siyuan Liu }}^{2}$ , Xu Huang 9, Yi-Zhou Sun ${}^{1}$ , Yuxiang Wang ${}^{7,8}$ , Zexi Zhang ${}^{5}$ , Zhunzhun Yu ${}^{3}$ , Sanzhong Luo ${}^{2 * }$ , Xue-Feng Fu ${}^{{10} * }$ & Shu-Li You ${}^{{11} * }$

洪鑫${}^{1 * }$ ，杨骐${}^{2 * }$ ，廖矿标${}^{3 * }$ ，裴剑锋${}^{4 * }$ ，陈茂${}^{5 * }$ ，莫凡洋${}^{6,8 * }$ ，吕华${}^{7 * }$ ，张文彬${}^{7,8 * }$ ，周海森${}^{7}$ ，陈家晓${}^{4 * }$ ，苏乐斌${}^{3}$ ，张硕卿${}^{1}$ ，${\text{ Siyuan Liu }}^{2}$ ，黄旭9，孙逸舟${}^{1}$ ，王玉香${}^{7,8}$ ，张泽熙${}^{5}$ ，余谆谆${}^{3}$ ，罗三中${}^{2 * }$ ，付雪峰${}^{{10} * }$ & 游书力${}^{{11} * }$

${}^{1}$ Center of Chemistry for Frontier Technologies, Department of Chemistry, Zhejiang University, Hangzhou 310027, China; ${}^{2}$ Center of Basic Molecular Science, Department of Chemistry, Tsinghua University, Beijing 100084, China; ${}^{3}$ Guangzhou National Laboratory, Guangzhou 510005, China;

${}^{1}$ 浙江大学化学系前沿技术化学中心，杭州310027，中国；${}^{2}$ 清华大学化学系基础分子科学中心，北京100084，中国；${}^{3}$ 广州国家实验室，广州510005，中国；

${}^{4}$ Center for Quantitative Biology, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing 100871, China; ${}^{5}$ State Key Laboratory of Molecular Engineering of Polymers, Department of Macromolecular Science, Fudan University, Shanghai 200433, China;

${}^{4}$ 北京大学前沿交叉学科研究院定量生物学中心，北京100871，中国；${}^{5}$ 复旦大学高分子科学系聚合物分子工程国家重点实验室，上海200433，中国；

${}^{6}$ School of Materials Science and Engineering, Peking University, Beijing 100871, China;

${}^{6}$ 北京大学材料科学与工程学院，北京100871，中国；

${}^{7}$ Beijing National Laboratory for Molecular Sciences, Center for Soft Matter Science and Engineering, Key Laboratory of Polymer Chemistry

${}^{7}$ 北京分子科学国家实验室，软物质科学与工程中心，聚合物化学与物理教育部重点实验室

and Physics of Ministry of Education, College of Chemistry and Molecular Engineering, Peking University, Beijing 100871, China;

北京大学化学与分子工程学院，北京100871，中国；

${}^{8}$ Al for Science (AI4S)-Preferred Program, Shenzhen Graduate School, Peking University, Shenzhen 518055, China;

${}^{8}$ 北京大学深圳研究生院人工智能促进科学(AI4S)优先项目，深圳518055，中国；

${}^{9}$ State Key Laboratory of Chemical Biology, Shanghai Institute of Organic Chemistry, University of Chinese Academy of Sciences, Chinese Academy of Sciences, Shanghai 200032, China;

${}^{9}$ 中国科学院上海有机化学研究所化学生物学国家重点实验室，中国科学院大学，上海200032，中国；

${}^{10}$ Department of Chemical Sciences, National Natural Science Foundation of China, Beijing 100085, China;

${}^{10}$ 中国化学科学部，国家自然科学基金委员会，北京10085，中国；

${}^{11}$ State Key Laboratory of Organometallic Chemistry, Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, Shanghai 200032, China

${}^{11}$ 中国科学院上海有机化学研究所金属有机化学国家重点实验室，上海200032，中国

Received March 20, 2024; accepted April 28, 2024; published online June 26, 2024

2024年3月20日收到；2024年4月28日接受；2024年6月26日在线发表

Recent years have witnessed the transformative impact from the integration of artificial intelligence with organic and polymer synthesis. This synergy offers innovative and intelligent solutions to a range of classic problems in synthetic chemistry. These exciting advancements include the prediction of molecular property, multi-step retrosynthetic pathway planning, elucidation of the structure-performance relationship of single-step transformation, establishment of the quantitative linkage between polymer structures and their functions, design and optimization of polymerization process, prediction of the structure and sequence of biological macromolecules, as well as automated and intelligent synthesis platforms. Chemists can now explore synthetic chemistry with unprecedented precision and efficiency, creating novel reactions, catalysts, and polymer materials under the data-driven paradigm. Despite these thrilling developments, the field of artificial intelligence (AI) synthetic chemistry is still in its infancy, facing challenges and limitations in terms of data openness, model interpretability, as well as software and hardware support. This review aims to provide an overview of the current progress, key challenges, and future development suggestions in the interdisciplinary field between AI and synthetic chemistry. It is hoped that this overview will offer readers a comprehensive understanding of this emerging field, inspiring and promoting further scientific research and development.

近年来，人工智能与有机和聚合物合成的融合产生了变革性影响。这种协同作用为合成化学中的一系列经典问题提供了创新和智能的解决方案。这些令人兴奋的进展包括分子性质预测、多步逆合成途径规划、单步转化的结构-性能关系阐明、聚合物结构与其功能之间定量联系的建立、聚合过程的设计和优化、生物大分子结构和序列的预测，以及自动化和智能合成平台。化学家现在可以以前所未有的精度和效率探索合成化学，在数据驱动的范式下创造新的反应、催化剂和聚合物材料。尽管有这些令人兴奋的发展，但人工智能(AI)合成化学领域仍处于起步阶段，在数据开放性、模型可解释性以及软件和硬件支持方面面临挑战和限制。本综述旨在概述人工智能与合成化学交叉领域的当前进展、关键挑战和未来发展建议。希望这一概述能让读者全面了解这一新兴领域，激发和促进进一步的科学研究与发展。

organic synthesis, polymer synthesis, machine learning prediction, chemical database, automated synthesis

有机合成、聚合物合成、机器学习预测、化学数据库、自动化合成

Citation: Hong X, Yang Q, Liao K, Pei J, Chen M, Mo F, Lu H, Zhang WB, Zhou H, Chen J, Su L, Zhang SQ, Liu S, Huang X, Sun YZ, Wang Y, Zhang Z, Yu Z, Luo S, Fu XF, You SL. AI for organic and polymer synthesis. Sci China Chem, 2024, 67: 2461-2496, https://doi.org/10.1007/s11426-024-2072-4

引用:Hong X, Yang Q, Liao K, Pei J, Chen M, Mo F, Lu H, Zhang WB, Zhou H, Chen J, Su L, Zhang SQ, Liu S, Huang X, Sun YZ, Wang Y, Zhang Z, Yu Z, Luo S, Fu XF, You SL. 用于有机和聚合物合成的人工智能。《中国科学:化学》，2024年，67卷:2461 - 2496页，https://doi.org/10.1007/s11426-024-2072-4

*Corresponding authors (; ; ; ; ; ; ; ; ; ; )

*通讯作者(邮箱:；；；；；；；；；；)

CONTENTS

																		1 Introduction 																			 2462

																		2 Machine learning pipeline 																			 2462

																		3 AI applications in organic synthesis 																			 2465

																			3.1 Molecular property prediction 																			 2465

																					3.2 Prediction and optimization of synthetic transfor- 		 mation 																			 2467

																		4 AI applications in polymer synthesis 																			 2475

																					4.1 Structure-property relationship prediction of poly- 		 mer 																			 2475

4.2 Target-orientated design of polymer

4.3 Design and optimization of polymer synthesis

																			4.4 End-to-end prediction of polymerization 																			 2479

																			4.5 AI Application in biological macromolecules 																			 2481

																		5 Automated experimentation 																			 2483

																			5.1 Automated synthesis 																			 2483

																				5.2 Automated work-up, isolation and purification 																				 2486

																			5.3 Integration of AI with robotic systems 																			 2486

																			6 Challenges and perspective 																				 2486

																				6.1 Data 																				 2486

																				6.2 Encoding 																				 2487

																				6.3 Model availability 																				 2488

																				6.4 Automated experimentation 																				 2488

																			7 Conclusions and outlook 																				 2489

1 Introduction

1 引言

Artificial intelligence (AI) encompasses a broad set of technologies that simulate human intelligence, of which machine learning (ML) is a crucial subset. ML enables computer systems to learn from and interpret data without explicit programming, forming the core mechanism behind most AI applications. By observing and analyzing massive datasets, AI algorithms can identify patterns, classify information, and even make complex decisions. Particularly in the field of natural language processing (NLP), the development of AI has been transformative. Recent years have seen large language models (LLMs) [1-4], like ChatGPT [5], significantly contribute to this advancement, embodying the dream of artificial general intelligence. In the field of chemistry, LLMs also have exciting applications: research has shown that LLMs inherently possess a certain degree of chemical understanding [6]. Models like Coscientist, which can autonomously design, plan, and execute chemical research, illustrate how LLMs facilitate chemical research by automating literature analysis and experimental processes [7]. This AI wave continues to expand, driving technological advancements across a broad spectrum of domains; for example, the advent of AlphaGo has seen it defeat top human players in Go [8], and the emergence of AlphaFold [9] signals the inevitable embrace of the AI revolution in natural sciences. This paradigm shift brought about by AI is profoundly influencing the methods through which humanity tackles and resolves complex high-dimensional problems.

人工智能(AI)涵盖了一系列模拟人类智能的技术，其中机器学习(ML)是一个关键子集。机器学习使计算机系统能够在无需明确编程的情况下从数据中学习并进行解释，构成了大多数人工智能应用背后的核心机制。通过观察和分析海量数据集，人工智能算法可以识别模式、分类信息，甚至做出复杂决策。特别是在自然语言处理(NLP)领域，人工智能的发展具有变革性。近年来，像ChatGPT [5]这样的大语言模型(LLMs)[1 - 4]对这一进展做出了重大贡献，体现了通用人工智能的梦想。在化学领域，大语言模型也有令人兴奋的应用:研究表明，大语言模型本身具有一定程度的化学理解能力[6]。像Coscientist这样能够自主设计、规划和执行化学研究的模型，说明了大语言模型如何通过自动化文献分析和实验过程来促进化学研究[7]。这股人工智能浪潮不断扩展，推动了广泛领域的技术进步；例如，AlphaGo的出现使其在围棋比赛中击败了顶尖人类棋手[8]，而AlphaFold [9]的出现标志着自然科学不可避免地迎来了人工智能革命。人工智能带来的这种范式转变正在深刻影响人类处理和解决复杂高维问题的方法。

Within the realm of synthetic chemistry, chemists are constantly faced with the challenges of complexity and multidimensionality. The inherent intricacy of these problems renders bottom-up theoretical deductions difficult, leading synthetic chemists to realize the potential of approaching these issues from the perspectives of data science and information science [10]. Whether in synthetic pathway planning [11-13] or exploring substituent effects [14], chemists have already widely applied data-driven methods. These methods, ranging from simple linear fitting to the development of complex expert systems, offered a powerful strategy for chemists to find solutions in the ocean of data, yielding fruitful advances in synthetic chemistry.

在合成化学领域，化学家们不断面临复杂性和多维度性的挑战。这些问题固有的复杂性使得自下而上的理论推导变得困难，促使合成化学家认识到从数据科学和信息科学的角度解决这些问题的潜力[10]。无论是在合成路线规划[11 - 13]还是探索取代基效应[14]方面，化学家们已经广泛应用了数据驱动的方法。这些方法从简单的线性拟合到复杂专家系统的开发，为化学家在数据海洋中寻找解决方案提供了强大的策略，在合成化学中取得了丰硕的进展。

From the research journey of the substituent effect, we can appreciate the profound impact of data and intelligence on synthetic chemistry. The pioneering explorations of Ingold [15] and Robinson et al. [16] laid the foundation for concepts such as steric hindrance and electronic effects, now fundamental in organic chemistry textbooks. Hammett’s systematic and in-depth application of linear relationship to the study of substituent effects has made the Hammett equation a cornerstone for analyzing organic reaction mechanisms [17,18]. Later, the exciting advances from Sigman and others revealed the potential of multivariate linear free energy relationship (LFER) in propelling the understanding and design of modern synthetic transformations [19]. Today, the continuous expansion of synthetic chemistry databases and the advancement of AI algorithms have enabled chemists to make chemical accuracy-level predictions of molecular properties directly from topological structures. A notable example includes the application of the $i$ BonD database [20] for $\mathrm{p}{K}_{\mathrm{a}}$ predictions, which matches the accuracy of quantum calculations while significantly enhancing efficiency by orders of magnitude [21]. These advancements have not only facilitated progress in organic synthesis, but also led to significant achievements in polymer synthesis and automated experimentation with important directions highlighted in Figure 1, signaling the dawn of a new era of intelligent synthesis. In this review, we focus not only on representative directions and outcomes of the intersection between artificial intelligence and synthetic chemistry, but also delve into the current challenges facing the field along with potential solutions. This list is by no means comprehensive, but it is our hope that through this review, readers will gain a clear and comprehensive perspective on the breadth and depth of AI applications in synthetic chemistry.

从取代基效应的研究历程中，我们可以领略到数据和智能对合成化学的深远影响。英戈尔德[15]以及罗宾逊等人[16]的开创性探索为空间位阻和电子效应等概念奠定了基础，这些概念如今已成为有机化学教科书的基石。哈梅特对线性关系在取代基效应研究中的系统而深入的应用，使哈梅特方程成为分析有机反应机理的基石[17,18]。后来，西格曼等人的令人兴奋的进展揭示了多元线性自由能关系(LFER)在推动现代合成转化的理解和设计方面的潜力[19]。如今，合成化学数据库的不断扩展和人工智能算法的进步使化学家能够直接从拓扑结构对分子性质进行化学精度级别的预测。一个显著的例子包括应用$i$邦德数据库[20]进行$\mathrm{p}{K}_{\mathrm{a}}$预测，其与量子计算的精度相当，同时效率提高了几个数量级[21]。这些进展不仅促进了有机合成的进步，也在聚合物合成和自动化实验方面取得了重大成就，图1突出了重要方向，标志着智能合成新时代的曙光。在本综述中，我们不仅关注人工智能与合成化学交叉领域的代表性方向和成果，还深入探讨该领域目前面临的挑战以及潜在的解决方案。这份清单绝非详尽无遗，但我们希望通过本综述，读者能对人工智能在合成化学中的应用广度和深度有清晰而全面的认识。

2 Machine learning pipeline

2机器学习流程

Prior to delving into specific research advancements, it is essential to elucidate the fundamentals of ML, especially as they pertain to applications within synthetic chemistry. ML techniques are categorized into three primary types: supervised learning, where the goal is to learn a function mapping inputs to outputs given labeled data; unsupervised learning, aimed at uncovering the hidden structure of unlabeled data; and reinforcement learning, focused on learning how to take actions to maximize some notion of cumulative reward through interaction with an environment. ML typically encompasses four critical stages: data collection, encoding, model training, and result analysis (Figure 2). Initially, the collection and organization of relevant data lay the groundwork for model construction. Subsequently, during the encoding phase, these data are transformed into a format interpretable by ML models. The model training stage then utilizes encoded data, allowing algorithms to identify patterns and relationships within the data. Finally, result analysis evaluates the predictive performance of the model as well as interprets the rationale behind the model predictions. These stages collectively form the foundation for applying ML in the realm of synthetic chemistry, promoting the intelligent solution of chemical problems.

在深入探讨具体的研究进展之前，有必要阐明机器学习的基本原理，特别是与合成化学中的应用相关的原理。机器学习技术主要分为三种类型:监督学习，目标是在给定标记数据的情况下学习将输入映射到输出的函数；无监督学习，旨在揭示未标记数据的隐藏结构；强化学习，专注于学习如何通过与环境交互采取行动以最大化某种累积奖励的概念。机器学习通常包括四个关键阶段:数据收集、编码、模型训练和结果分析(图2)。首先，相关数据的收集和整理为模型构建奠定了基础。随后，在编码阶段，这些数据被转换为机器学习模型可解释的格式。然后，模型训练阶段利用编码后的数据，使算法能够识别数据中的模式和关系。最后，结果分析评估模型的预测性能，并解释模型预测背后的原理。这些阶段共同构成了在合成化学领域应用机器学习的基础，促进了化学问题的智能解决。

The primary sources of data in synthetic chemistry currently include public databases, high-throughput experimentation (HTE), computational simulations, and electronic laboratory notebooks (ELNs). These diverse data streams are vital for the success of ML modeling, offering extensive information on reactions, compounds, and properties. Table 1 lists exemplary open-access databases for organic and polymer synthesis. Public databases like Reaxys and Sci-finder are indispensable for providing comprehensive chemical data, while HTE systems enable the efficient generation of large datasets through automated experiments, assessing thousands of reactions with minimal material and time. ELNs play a crucial role in documenting and sharing experimental details [22,23], although they present challenges related to data standardization, privacy, and variability. Together, these sources underpin the development of ML models in synthetic chemistry, leveraging the vast array of data to fuel innovations through computational analysis and experimental integration.

合成化学中数据的主要来源目前包括公共数据库、高通量实验(HTE)、计算模拟和电子实验室笔记本(ELN)。这些不同的数据流对于机器学习建模的成功至关重要，提供了有关反应、化合物和性质的广泛信息。表1列出了有机和聚合物合成的示例性开放获取数据库。像Reaxys和Sci-finder这样的公共数据库对于提供全面的化学数据不可或缺，而高通量实验系统通过自动化实验能够高效生成大型数据集，以最少的材料和时间评估数千个反应。电子实验室笔记本在记录和共享实验细节方面发挥着关键作用[22,23]，尽管它们在数据标准化、隐私和可变性方面存在挑战。总之，这些来源支撑着合成化学中机器学习模型的发展，利用大量数据通过计算分析和实验整合推动创新。

Encoding molecules and reactions into machine-readable formats are critical for ML modeling. The molecular encodings can be characterized by a hierarchy of complexity: zero-dimensional physicochemical properties like molecular weight and $\operatorname{Log}P$ , one-dimensional string representations such as SMILES [38] and SELFIES [39] for encoding atomic types and connections, two-dimensional molecular fingerprints capturing molecular structures without stereochemical details, and three-dimensional descriptors that include stereochemistry and quantum chemical features for a comprehensive representation of molecular conformations. Additionally, graph-based learning methods offer advanced ways to depict molecules and reactions [40], addressing the complexity of chemical structures in multidimensional space. For polymers, the challenge of their stochastic nature is met with novel encoding strategies like BigSMILES [41] for sequence distributions and PolyGrammar [42] for hy-pergraph representations, alongside graph neural networks and Transformer-based language models to distinguish polymer sequences and topologies. These encoding strategies are essential for effectively processing and analyzing chemical data, enabling the advancement of ML applications in organic and polymer systems.

将分子和反应编码为机器可读格式对于机器学习建模至关重要。分子编码可以按复杂度层次进行分类:零维物理化学性质，如分子量和$\operatorname{Log}P$，一维字符串表示，如用于编码原子类型和连接的SMILES [38]和SELFIES [39]，二维分子指纹，用于捕获无立体化学细节的分子结构，以及三维描述符，包括立体化学和量子化学特征，用于全面表示分子构象。此外，基于图的学习方法提供了描绘分子和反应的先进方式[40]，解决了多维空间中化学结构的复杂性。对于聚合物，其随机性质的挑战通过新的编码策略来应对，如用于序列分布的BigSMILES [41]和用于超图表示的PolyGrammar [42]，以及图神经网络和基于Transformer的语言模型来区分聚合物序列和拓扑结构。这些编码策略对于有效处理和分析化学数据至关重要，推动了机器学习在有机和聚合物系统中的应用发展。

Figure 1 Representative research directions of AI applications in organic and polymer synthesis (color online).

图1人工智能在有机和聚合物合成中的代表性研究方向(在线彩色图)。

Figure 2 Typical pipeline of machine learning modeling (color online).

图2机器学习建模的典型流程(在线彩色图)。

Table 1 Overview of representative open access databases for organic and polymer synthesis

表1有机和聚合物合成的代表性开放获取数据库概述

The process of ML modeling involves utilizing algorithms to grasp patterns within data, thereby enabling predictions about unknown targets. This involves a spectrum of methodologies from classical ML, adept at handling linear relationships and structured data, to deep learning, known for its proficiency with large-scale and complex datasets. Classical models like linear, tree-based, and kernel-based methods offer solutions for simpler relationships, while deep learning’s layered architecture allows for the extraction of high-dimensional features. In addition, the modern ML process can be adaptive with active learning and transfer learning techniques. Active learning dynamically selects the most informative data points for labeling and training, effectively improving model performance with less data. Transfer learning leverages knowledge acquired from one domain to enhance model accuracy in another, significantly reducing the need for extensive labeled datasets in new applications. Selecting the right model and algorithm is crucial, depending on the data’s nature and the analytical task at hand. Typical evaluation methods include cross-validation and independent test sets, which are essential to ensure the model’s generalizability and effectiveness in real-world applications.

机器学习建模过程涉及利用算法掌握数据中的模式，从而能够对未知目标进行预测。这涉及从擅长处理线性关系和结构化数据的经典机器学习到以处理大规模和复杂数据集而闻名的深度学习等一系列方法。像线性方法、基于树的方法和基于核的方法这样的经典模型为更简单的关系提供解决方案，而深度学习的分层架构允许提取高维特征。此外，现代机器学习过程可以通过主动学习和迁移学习技术实现自适应。主动学习动态选择最具信息性的数据点进行标记和训练，用更少的数据有效提高模型性能。迁移学习利用从一个领域获得的知识来提高另一个领域的模型准确性，显著减少新应用中对大量标记数据集的需求。根据数据的性质和手头的分析任务选择合适的模型和算法至关重要。典型的评估方法包括交叉验证和独立测试集，这对于确保模型在实际应用中的通用性和有效性至关重要。

For ML applications, result analysis and model interpretation are equally important [43,44], as merely relying on and executing each model prediction is insufficient and lacks comprehensiveness. In the analysis of ML predictions, one can leverage domain expertise to evaluate and contrast predictions against non-ML methodologies. Techniques such as sensitivity analysis and hypothesis testing further aid in assessing the accuracy and reliability of models, especially when predictions deviate from expected norms. Additionally, methods like dimensionality reduction and clustering are instrumental in deriving valuable insights from ML predictions, as demonstrated in Tim Cernak’s utilization of graph editing distance to analyze retrosynthesis routes designed by SYNTHIA [45]. The goal of model interpretation is to elucidate the decision-making process of complex, high-dimensional ML models, thereby extracting heuristic principles and knowledge pertinent to the domain. Crucial approaches include feature importance analysis, which pinpoints key influencing variables, and interpretability frameworks such as SHAP [46,47] and LIME [48,49], which facilitate the understanding of how models arrive at their decisions.

对于机器学习应用而言，结果分析和模型解释同样重要[43,44]，因为仅仅依赖并执行每个模型预测是不够的，且缺乏全面性。在机器学习预测分析中，可以利用领域专业知识来评估预测结果，并与非机器学习方法进行对比。诸如敏感性分析和假设检验等技术有助于进一步评估模型的准确性和可靠性，尤其是当预测结果偏离预期标准时。此外，降维和聚类等方法有助于从机器学习预测中获得有价值的见解，正如蒂姆·切尔纳克利用图编辑距离分析SYNTHIA设计的逆合成路线所展示的那样[45]。模型解释的目标是阐明复杂的高维机器学习模型的决策过程，从而提取与该领域相关的启发式原则和知识。关键方法包括特征重要性分析，它能找出关键影响变量，以及诸如SHAP[46,47]和LIME[48,49]等可解释性框架，这些框架有助于理解模型是如何做出决策的。

3 AI applications in organic synthesis

3人工智能在有机合成中的应用

In the realm of organic synthesis, data-driven methodologies are catalyzing solutions to a multitude of complex challenges, achieving substantial progress in recent years. These issues encompass a vast array of dimensions, from single molecule to intermolecular chemical reactions, and even multi-step transformations in total synthesis. The scope of research in this area is equally comprehensive, involving the prediction of molecular physicochemical properties, the evaluation of structure-activity relationships in organic transformations, and the optimization of reaction conditions. Facing these chemical challenges across diverse scenarios, researchers have applied and developed innovative AI technologies, with key directions highlighted in Figure 3. These successes demonstrate the immense potential of AI within the sphere of organic synthesis. In this section, the representative research advances are discussed to showcase the ability of AI technology in these application scenarios.

在有机合成领域，数据驱动的方法正在催生众多复杂挑战的解决方案，近年来取得了重大进展。这些问题涵盖了广泛的维度，从单分子到分子间化学反应，甚至全合成中的多步转化。该领域的研究范围同样广泛，包括分子物理化学性质的预测、有机转化中构效关系的评估以及反应条件的优化。面对各种场景下的这些化学挑战，研究人员应用并开发了创新的人工智能技术，关键方向如图3所示。这些成功证明了人工智能在有机合成领域的巨大潜力。在本节中，将讨论代表性的研究进展，以展示人工智能技术在这些应用场景中的能力。

3.1 Molecular property prediction

3.1 分子性质预测

The physicochemical properties of organic molecules dictate how they behave in chemical reactions and influence the evolution of organic transformations [50,51]. Therefore, the accurate comprehension and prediction of molecular properties serve as the backbone for the rationality of organic synthesis. As shown in Table 2, molecular properties can be classified into thermodynamic and kinetic parameters, with data sourced from both experiment and computation. Through the discussion of the highlighted ML studies, this section will elaborate on how ML enables quantitative thermodynamic and kinetic property predictions.

有机分子的物理化学性质决定了它们在化学反应中的行为方式，并影响有机转化的进程[50,51]。因此，准确理解和预测分子性质是有机合成合理性的基础。如表2所示，分子性质可分为热力学和动力学参数，数据来源于实验和计算。通过对重点机器学习研究的讨论，本节将阐述机器学习如何实现热力学和动力学性质的定量预测。

(1) Prediction of thermodynamic properties

(1) 热力学性质的预测

Thermodynamic properties are inherent characteristics of molecules in an equilibrium state that are used as fundamental parameters to assess the thermodynamics of chemical reactions. Traditional approaches for determining thermodynamic parameters entail both experimental methods and quantum chemical computations, which are accurate but also time- and resource-intensive [110,111]. However, since the thermodynamic property is dictated by the molecular structure, it presents a scenario that is well-suited for ML modeling and prediction. This high-dimensional mapping from molecular structure to thermodynamic property can be learned by data-driven approaches, which could lead to much more accurate and efficient thermodynamic parameter prediction than with conventional techniques. In recent years, significant breakthroughs have been seen in the prediction of various important thermodynamic parameters including $\mathrm{p}{K}_{\mathrm{a}}$ , BDE, and others.

热力学性质是分子在平衡状态下的固有特性，用作评估化学反应热力学的基本参数。传统的确定热力学参数的方法需要实验方法和量子化学计算，这些方法虽然准确，但也耗费时间和资源[110,111]。然而，由于热力学性质由分子结构决定，这为机器学习建模和预测提供了一个合适的场景。从分子结构到热力学性质的这种高维映射可以通过数据驱动的方法来学习，这可能比传统技术带来更准确、更高效的热力学参数预测。近年来，在包括$\mathrm{p}{K}_{\mathrm{a}}$、键解离能(BDE)等各种重要热力学参数预测方面取得了重大突破。

$\mathrm{p}{K}_{\mathrm{a}}$ indicates the degree of proton dissociation from a molecule, which is important to understanding heterolytic X-H bond cleavage energies [112-116] and plays a critical role in both chemical and medical sciences [57,114-116]. Although quantum mechanical computations have been extensively applied for $\mathrm{p}{K}_{\mathrm{a}}$ evaluation with high accuracy [54- 58], they also suffer from time- and resource-consuming limitations. To achieve the data-driven $\mathrm{p}{K}_{\mathrm{a}}$ prediction, classic ML methods [21,57-65] and graph convolutional neural networks (GCNs) based approaches [66,67] have made tremendous progress. Benefiting from the massive $\mathrm{p}{K}_{\mathrm{a}}$ values in the $i$ BonD database, Luo et al. [21] have reported an NN-based ML model that can predict the overall $\mathrm{p}{K}_{\mathrm{a}}$ value of a given molecule (macro-p ${K}_{\mathrm{a}}$ ) with an MAE of ${0.87}\mathrm{p}{K}_{\mathrm{a}}$ units in various solvents [20]. For micro-p ${K}_{\mathrm{a}}$ of a specific X-H bond, Grzybowski et al. [66] achieved a mean absolute error (MAE) of ${2.1}\mathrm{p}{K}_{\mathrm{a}}$ units using a GCN model with a DFT-calculated database, which enabled accurate $\mathrm{p}{K}_{\mathrm{a}}$ prediction of a wide range of $\mathrm{C} - \mathrm{H}$ acids. These strategies have also been extended to the $\mathrm{p}{K}_{\mathrm{a}}$ prediction in protein residues [117,118], which play a crucial role in regulating protein structures and their functions in biological processes.

$\mathrm{p}{K}_{\mathrm{a}}$表示分子中质子解离的程度，这对于理解异裂X-H键的裂解能很重要[112 - 116]，并且在化学和医学科学中都起着关键作用[57,114 - 116]。尽管量子力学计算已被广泛应用于高精度的$\mathrm{p}{K}_{\mathrm{a}}$评估[54 - 58]，但它们也存在耗时和资源消耗大的局限性。为了实现数据驱动的$\mathrm{p}{K}_{\mathrm{a}}$预测，经典机器学习方法[21,57 - 65]和基于图卷积神经网络(GCN)的方法[66,67]取得了巨大进展。受益于$i$化学键数据库中的大量$\mathrm{p}{K}_{\mathrm{a}}$值，Luo等人[21]报告了一种基于神经网络的机器学习模型，该模型可以预测给定分子在各种溶剂中的整体$\mathrm{p}{K}_{\mathrm{a}}$值(宏观p${K}_{\mathrm{a}}$)，平均绝对误差(MAE)为${0.87}\mathrm{p}{K}_{\mathrm{a}}$单位[20]。对于特定X-H键的微观p${K}_{\mathrm{a}}$，Grzybowski等人[66]使用具有密度泛函理论(DFT)计算数据库的GCN模型实现了${2.1}\mathrm{p}{K}_{\mathrm{a}}$单位的平均绝对误差(MAE)，这使得能够对广泛的$\mathrm{C} - \mathrm{H}$酸进行准确的$\mathrm{p}{K}_{\mathrm{a}}$预测。这些策略也已扩展到蛋白质残基中的$\mathrm{p}{K}_{\mathrm{a}}$预测[117,118]，这在调节蛋白质结构及其在生物过程中的功能方面起着至关重要的作用。

Figure 3 Key directions of AI applications in organic synthesis (color online).

图3 人工智能在有机合成中的应用关键方向(彩色在线)。

Table 2 Overview of representative molecular properties and models

表2 代表性分子性质和模型概述

BDE, which involves the homolysis of chemical bonds, reflects the intrinsic bond strength and is critical in a series of chemical transformations. One representative example is the metal-oxo complex-mediated C-H activation [114], in which the C-H BDEs are closely related to reaction rates. Typically, BDEs could be determined using experimental methods [119] or theoretical calculations with an MAE of around 2 kcal/mol [68-70]. However, these methods, while precise, are costly and inefficient for large-scale analysis. Over the last two decades, early QSAR studies laid the groundwork for efficient and accurate approaches to evaluating BDEs [69-75]. More recent advancements leveraged high-throughput DFT calculations to generate larger BDE datasets with diverse bond types and advanced ML strategies like GNNs [76,77] and spectrum-enhanced strategies [78], offering high-accuracy BDE predictions [79-82]. It is noted that Paton et al. [81] have reported an appealing GNN model based on approximately 300k DFT-calculated BDEs, achieving accuracy with an MAE of 0.58 kcal/mol when compared with DFT calculations.

键解离能(BDE)涉及化学键的均裂，反映了内在键强度，并且在一系列化学转化中至关重要。一个典型的例子是金属-氧配合物介导的C-H活化[114]，其中C-H键解离能与反应速率密切相关。通常，键解离能可以使用实验方法[119]或理论计算来确定，平均绝对误差约为2千卡/摩尔[68 - 70]。然而，这些方法虽然精确，但对于大规模分析来说成本高且效率低。在过去二十年中，早期的定量构效关系(QSAR)研究为评估键解离能的高效准确方法奠定了基础[69 - 75]。最近的进展利用高通量密度泛函理论计算来生成具有不同键类型的更大键解离能数据集，并采用诸如图神经网络(GNN)[76,77]和光谱增强策略[78]等先进的机器学习策略，提供高精度的键解离能预测[79 - 82]。值得注意的是，Paton等人[81]报告了一种基于约300k密度泛函理论计算的键解离能的有吸引力的图神经网络模型，与密度泛函理论计算相比，平均绝对误差为0.58千卡/摩尔时达到了高精度。

In addition to $\mathrm{p}{K}_{\mathrm{a}}$ and BDE, the ML modeling of a collection of quantum chemical properties (HOMO/LUMO energies, thermal values like $U, H, G$ , etc.) of molecules has made significant progress thanks to the creation of the quantum machine (QM)-series database [36,37,120-122]. This large-scale database serves as a powerful data engine that stimulated the development of a series of novel AI frameworks for molecular prediction, including SchNet [40], PhysNet [83], HMGNN [84], TensorNet [85], ChemRL-GEM [86], and Uni-Mol [87]. These models continued to push the state-of-the-art (SOTA) record of molecular property prediction in the QM-series database, achieving an accuracy comparable to DFT calculations. Aside from the synthetic interest, there has been a long-standing interest in predicting molecular properties that are important for drug and material design. Standard databases such as ESOL [88], FreeSolv [89], Solv@TUM [90], and Lipophilicity [91] have been widely used in developing accurate predictive models to aid in the design and screening of drug-like molecules. Redox potential, on the other hand, is an important parameter in electrochemical behavior and has been modeled using a variety of molecular representations, including HOMO/ LUMO orbital energy [92] and DFT-calculated descriptors [93].

除了$\mathrm{p}{K}_{\mathrm{a}}$和键解离能(BDE)之外，由于量子机器(QM)系列数据库的创建[36,37,120 - 122]，分子的一系列量子化学性质(最高占据分子轨道/最低未占据分子轨道能量、像$U, H, G$这样的热学值等)的机器学习建模取得了显著进展。这个大规模数据库作为一个强大的数据引擎，推动了一系列用于分子预测的新型人工智能框架的发展，包括SchNet [40]、PhysNet [83]、HMGNN [84]、TensorNet [85]、ChemRL - GEM [86]和Uni - Mol [87]。这些模型不断刷新QM系列数据库中分子性质预测的当前最优(SOTA)记录，达到了与密度泛函理论(DFT)计算相当的精度。除了合成方面的兴趣，长期以来人们一直对预测药物和材料设计中重要的分子性质感兴趣。诸如ESOL [88]、FreeSolv [89]、Solv@TUM [90]和亲脂性[91]等标准数据库已被广泛用于开发准确的预测模型，以协助类药物分子的设计和筛选。另一方面，氧化还原电位是电化学行为中的一个重要参数，已使用多种分子表示进行建模，包括最高占据分子轨道/最低未占据分子轨道能量[92]和DFT计算的描述符[93]。

(2) Prediction of kinetic properties

(2) 动力学性质预测

In synthetic chemistry, kinetic properties are just as important as thermodynamic properties in determining the practical feasibility of reactions that are theoretically favorable. Their importance extends to critical aspects such as reaction yield, regioselectivity, and stereoselectivity, and they are essential in a variety of processes, including pharmacokinetics [123,124], dynamic kinetic resolution [125], and petroleum cracking [126]. However, the development of kinetic property prediction lags behind that of thermodynamic properties due to the intrinsic complexity and limited data availability [70,127,128].

在合成化学中，动力学性质在确定理论上可行的反应的实际可行性方面与热力学性质同样重要。它们的重要性延伸到诸如反应产率、区域选择性和立体选择性等关键方面，并且在包括药代动力学[123,124]、动态动力学拆分[125]和石油裂解[126]在内的各种过程中至关重要。然而，由于内在的复杂性和有限的数据可用性[70,127,128]，动力学性质预测的发展落后于热力学性质预测。

Reaction rate constants and activation energies are key kinetic properties worthy of ML modeling [129]. In 2020, Green et al. [94] utilized a message-passing neural network model with reaction fingerprints to predict activation energies, achieving an MAE of 1.7 kcal/mol. Using transition state modeling, Buttar et al. [95] predicted the barrier of nucleophilic aromatic substitution processes with an MAE of 0.77 kcal/mol. For rate constant prediction, Bowman and colleagues [96] employed Gaussian process regression for rate constant prediction for bimolecular chemical reactions. Greaves et al. [130] reported a multiple linear regression method to predict the rate constant of the reaction between benzyl bromide and pyridine with an ${R}^{2}$ of 0.92 . However, the lack of a substantial rate constant database limits the scope and application of these ML models, highlighting an important subject for future research and data collection efforts.

反应速率常数和活化能是值得进行机器学习建模的关键动力学性质[129]。2020年，格林等人[94]利用带有反应指纹的消息传递神经网络模型预测活化能，平均绝对误差(MAE)为1.7千卡/摩尔。使用过渡态建模，布塔尔等人[95]预测亲核芳香取代过程的能垒，MAE为0.77千卡/摩尔。对于速率常数预测，鲍曼及其同事[96]采用高斯过程回归预测双分子化学反应的速率常数。格里夫斯等人[130]报告了一种多元线性回归方法，用于预测苄基溴与吡啶反应的速率常数，${R}^{2}$为0.92。然而，缺乏大量的速率常数数据库限制了这些机器学习模型的范围和应用，凸显了未来研究和数据收集工作的一个重要课题。

As fundamental concepts in polar chemistry, nucleophilicity $\left( N\right)$ and electrophilicity $\left( E\right)$ are quantified with rate constants in specific reactions. Particularly, Mayr et al. established the well-known Mayr equation to describe the nucleophilicity and electrophilicity of molecules [131] and then built a database for $N/E$ evaluation in chemical reactions [132]. Several ML modeling attempts have been made based on the empirically known $N/E$ values to forecast the $N/E$ value of novel reagents using physical, topological, or quantum chemical descriptors [97-102] or directly via the GNN model [103]. Recently, Luo et al. [97] developed a holistic model for predicting both $N\left( {{R}^{2} = {0.92},\mathrm{{MAE}} = {1.45}}\right)$ and $\left( {{R}^{2} = {0.93},\mathrm{{MAE}} = {1.45}}\right)$ by integrating reactivity structural and physicochemical (rSPOC) descriptors, which was then used to predict the nucleophilicity of a variety of enamine intermediates and NAD(P)H.

作为极性化学中的基本概念，亲核性$\left( N\right)$和亲电性$\left( E\right)$在特定反应中用速率常数来量化。特别是，迈尔等人建立了著名的迈尔方程来描述分子的亲核性和亲电性[131]，然后建立了一个用于化学反应中$N/E$评估的数据[132]。基于经验已知的$N/E$值，已经进行了几次机器学习建模尝试，使用物理、拓扑或量子化学描述符[97 - 102]或直接通过图神经网络(GNN)模型[103]来预测新型试剂的$N/E$值。最近，罗等人[97]通过整合反应性结构和物理化学(rSPOC)描述符开发了一个用于预测$N\left( {{R}^{2} = {0.92},\mathrm{{MAE}} = {1.45}}\right)$和$\left( {{R}^{2} = {0.93},\mathrm{{MAE}} = {1.45}}\right)$的整体模型，然后用于预测各种烯胺中间体和烟酰胺腺嘌呤二核苷酸(磷酸)(NAD(P)H)的亲核性。

(3) Prediction of potential energy surface

(3) 势能面预测

Unlike the thermodynamic and kinetic properties, which are characterized by specific numerical values, ML modeling of potential energy surfaces (PESs) requires capturing the continuous relationship between nuclear coordinates and their corresponding energies. The precise prediction of PES is not only theoretically important but also has great practical significance. The AI potential model has received extensive attention in recent years [104,127,133-135], and it can help us understand and simulate molecular systems [136] and synthetic processes [137,138]. Behler and Parrinello made a seminal contribution to this field in 2007 when they used neural networks to create PES at remarkable speeds and efficiency [105]. This was followed by the introduction of Deep Potential Net in 2017 [106], which achieved quantum chemical precision in generating PESs. Further advancements include Schütt et al.’s use of SchNet for PES prediction and its application in molecular dynamics simulations of small molecules [40] as well as Clementi et al.’s development of CGnets for coarse-grained (CG) molecular modeling [107], extending it to encompass all-atom free energy surfaces in explicit solvation. In recent years, there has been a surge in exciting innovations, such as neural network-based full-dimensional PES constructions for chemical systems [104], Δ-machine learned PES enhancing DFT-based PESs to near CCSD(T) accuracy [108], and the enrichment of the PES library with more comprehensive datasets for chemical systems [139]. These developments have paved the way for an increasing number of modern methods exploring PESs of chemical reactions [140]. One representative example of using PES to elaborate synthetic mechanisms is Liu’s study of the mechanism and selectivity of glucose pyrolysis [109]. Using the AI potential model developed by their stochastic surface walking method [141-145], this study detailed an amount of 6,407 elementary reactions and elucidated the mechanistic details and origins of site-selectivity for 5-hy-droxymethylfurfural formation.

与以特定数值为特征的热力学和动力学性质不同，势能面(PES)的机器学习建模需要捕捉核坐标与其相应能量之间的连续关系。精确预测PES不仅在理论上很重要，而且具有重大的实际意义。近年来，人工智能势能模型受到了广泛关注[104,127,133 - 135]，它可以帮助我们理解和模拟分子系统[136]以及合成过程[137,138]。2007年，Behler和Parrinello在该领域做出了开创性贡献，他们使用神经网络以极高的速度和效率创建了PES[105]。随后在2017年引入了深度势能网络[106]，其在生成PES方面达到了量子化学精度。进一步的进展包括Schütt等人将SchNet用于PES预测及其在小分子分子动力学模拟中的应用[40]，以及Clementi等人开发用于粗粒度(CG)分子建模的CGnets[107]，并将其扩展到明确溶剂化中的全原子自由能表面。近年来，出现了大量令人兴奋的创新，例如用于化学系统的基于神经网络的全维PES构建[104]、将基于密度泛函理论(DFT)的PES增强到接近耦合簇单双激发加微扰三重激发(CCSD(T))精度的Δ机器学习PES[108]，以及用更全面的化学系统数据集丰富PES库[139]。这些发展为越来越多探索化学反应PES的现代方法铺平了道路[140]。使用PES阐述合成机制的一个代表性例子是Liu对葡萄糖热解机制和选择性的研究[109]。该研究使用他们的随机表面行走方法[141 - 145]开发的人工智能势能模型，详细阐述了6407个基元反应，并阐明了5 - 羟甲基糠醛形成的位点选择性的机理细节和起源。

3.2 Prediction and optimization of synthetic transfor- mation

3.2 合成转化的预测与优化

Because of the massive possibilities of chemical bond cleavage and formation, synthetic transformation inherently poses a multiple-choice question with numerous possible products. Moreover, the issue of synthetic transformation prediction also involves predicting the quantitative outcomes of reactions (yield, selectivity, etc.) and strategizing the synthetic pathways for multistep transformations. These problems are highly amenable to data-driven solutions. In fact, even before the advent of modern artificial intelligence technology, the birth and growth of chemoinformatics encompassed the exploration of using data and programming to address these challenges. In recent years, with the accumulation of large-scale synthetic data and the development of advanced reaction modeling frameworks, this field has seen significant progress. On a range of reaction prediction scenarios, AI has demonstrated promising prospects, even offering judgments that surpass those of human chemists.

由于化学键断裂和形成的可能性众多，合成转化本质上是一个有众多可能产物的选择题。此外，合成转化预测问题还涉及预测反应的定量结果(产率、选择性等)以及为多步转化制定合成途径策略。这些问题非常适合采用数据驱动的解决方案。事实上，甚至在现代人工智能技术出现之前，化学信息学的诞生和发展就涵盖了利用数据和编程来应对这些挑战的探索。近年来，随着大规模合成数据的积累和先进反应建模框架的发展，该领域取得了显著进展。在一系列反应预测场景中，人工智能已展现出广阔前景，甚至能做出超越人类化学家的判断。

(1) Multistep retrosynthesis planning

(1) 多步逆合成规划

Computer-assisted synthetic planning (CASP), particularly the strategic planning of multi-step retrosynthesis, represents one of the oldest yet most vibrant challenges in AI synthetic chemistry [146]. The crux of the challenge in ret-rosynthesis planning lies in the construction of a coherent and reasonable multi-step synthetic network, followed by the execution of an efficient and rational search and scoring process within this network. Finally, the synthesis pathways must be regressed to the available building blocks, ensuring a practical and feasible approach to the synthesis design. To realize the multistep retrosynthesis, a series of innovative algorithm designs have been proposed in recent years. In this regard, Segler et al. [147,148] utilized the Monte Carlo Tree Search (MCTS) algorithm to devise synthetic routes for small organic molecules. Kishimoto et al. [149] introduced the DFPN-E method, integrating depth-first proof-number search (DFPN) with heuristic edge initialization, showcasing a time advantage over the MCTS algorithm with comparable success rates. Chen et al. [150] presented Retro*, a neural-based A*-like algorithm, utilizing an AND-OR search tree and an optimal priority search strategy, offering a more efficient approach to searching reaction pathways. Xie et al. [151] presented a graph-based search algorithm called Ret-roGraph, further enhancing the performance of A*-like search algorithms to reduce molecular redundancy in tree-based search methods. Kim et al. [152] introduced Retro*+, a self-improving framework training a single-step model to emulate successful trajectories, maximizing success rates and leveraging simulated experiences for model enhancement. Yu et al. [153] proposed GRASP, a goal-driven actor-critic method, utilized for seeking routes with specific predefined objectives, such as building block materials. Recently, Liu et al. [154] presented PDVN, a dual-value network planning, constructing two distinct value networks to predict synthesizability and cost, enhancing search success rates, optimizing model invocations, and aiding in identifying shorter synthetic routes.

计算机辅助合成规划(CASP)，尤其是多步逆合成的战略规划，是人工智能合成化学中最古老但最具活力的挑战之一[146]。逆合成规划挑战的关键在于构建一个连贯且合理的多步合成网络，然后在该网络内执行高效且合理的搜索和评分过程。最后，合成途径必须回归到可用的构建模块，确保合成设计的实际可行性。为实现多步逆合成，近年来提出了一系列创新算法设计。在这方面，Segler等人[147,148]利用蒙特卡罗树搜索(MCTS)算法为小分子设计合成路线。Kishimoto等人[149]引入了DFPN-E方法，将深度优先证据数搜索(DFPN)与启发式边初始化相结合，在成功率相当的情况下，展示出比MCTS算法更具时间优势。Chen等人[150]提出了Retro*，一种基于神经网络的类似A的算法，利用与或搜索树和最优优先级搜索策略，提供了一种更高效的搜索反应途径的方法。Xie等人[151]提出了一种基于图的搜索算法Ret-roGraph，进一步提高了类似A搜索算法的性能，以减少基于树的搜索方法中的分子冗余。Kim等人[152]引入了Retro*+，一个自我改进框架，训练单步模型以模拟成功轨迹，最大化成功率并利用模拟经验增强模型。Yu等人[153]提出了GRASP，一种目标驱动的演员评论家方法，用于寻找具有特定预定义目标(如构建模块材料)的路线。最近，Liu等人[154]提出了PDVN，一种双值网络规划，构建两个不同的值网络来预测合成可行性和成本，提高搜索成功率，优化模型调用，并有助于识别更短的合成路线。

With the above algorithm advancements, reports on computer-aided multi-step route design in chemical synthesis are emerging. The Jensen group [155] employed ASKCOS for multi-step retrosynthetic route design of 15 drug molecules, including $\left( S\right)$ -warfarin and safinamide, and validated the synthetic feasibility through a robotic flow chemistry platform. The Grzybowski team [12] demonstrated SYNTHIA’s powerful retrosynthetic design capabilities, passing the Turing test in which chemists cannot differentiate the AI-designed and the human-designed synthetic routes for the studied compounds. A series of SYNTHIA-predicted routes for natural products are experimentally executed, including challenging targets of (-)-Dauricine, Tacamonidine, and Lamellodysidine A (Figure 4a). It is worth noticing that Tacamonidine and Lamellodysidine A were synthesized for the first time. The Cernak research group [156] utilized SYNTHIA to study 12 potential anti-COVID-19 drugs, experimentally validating four predicted routes for umifenovir and one predicted route for bromhexine, highlighting that automated retrosynthetic predictions can rapidly identify alternative starting material supply chains for pharmaceuticals. Cernak and colleagues [45] further demonstrated that human chemists can harness the heuristic value of AI predictions to achieve out-of-box synthetic innovations. They group-employed SYNTHIA for the retrosynthetic route prediction of (-)-stemoamide. Based on the myriad of SYNTHIA-predicted pathways, they proposed a concept of graph edit distance to quantify the synthetic impact of AI-suggested single-step transformations (Figure 4b). Through this, they were able to realize a remarkable 3-step synthesis of (-)-stemoamide (Figure 4c). Interestingly, the AI tool for retrosynthesis analysis can also be applied in a reversed fashion to guide forward synthetic possibilities. The Grzy-bowski team [157] utilized the forward-synthesis Allchemy platform to generate a plethora of synthetic networks from approximately 200 commercially recovered waste chemicals. They selected numerous viable synthetic routes and experimentally validated several of them. The continuous reporting of computer-aided multi-step reaction planning, encompassing both algorithm development and experimental applications, underscores the growing significance of this field in the work of synthetic chemists.

随着上述算法的进步，化学合成中计算机辅助多步路线设计的报告不断涌现。詹森团队[155]使用ASKCOS对15种药物分子进行多步逆合成路线设计，包括$\left( S\right)$ -华法林和沙芬酰胺，并通过机器人流动化学平台验证了合成可行性。格鲁兹鲍夫斯基团队[12]展示了SYNTHIA强大的逆合成设计能力，通过了图灵测试，在该测试中，化学家无法区分AI设计的和人工设计的所研究化合物的合成路线。一系列SYNTHIA预测的天然产物路线通过实验得以执行，包括具有挑战性的(-)-蝙蝠葛碱、他卡莫尼定和片螺素A的目标合成(图4a)。值得注意的是，他卡莫尼定和片螺素A是首次合成。切尔纳克研究小组[156]利用SYNTHIA研究了12种潜在的抗COVID-19药物，通过实验验证了乌米芬诺尔的4条预测路线和溴己新的1条预测路线，突出了自动化逆合成预测可以快速识别药物的替代起始原料供应链。切尔纳克及其同事[45]进一步证明，人类化学家可以利用AI预测的启发式价值实现突破性的合成创新。他们团队使用SYNTHIA对(-)-百部酰胺进行逆合成路线预测。基于SYNTHIA预测的众多途径，他们提出了图编辑距离的概念，以量化AI建议的单步转化的合成影响(图4b)。通过这种方法，他们能够实现显著的三步合成(-)-百部酰胺(图4c)。有趣的是，用于逆合成分析的AI工具也可以以相反的方式应用，以指导正向合成可能性。格鲁兹鲍夫斯基团队[157]利用正向合成Allchemy平台从大约200种商业回收的废化学品中生成了大量的合成网络。他们选择了许多可行的合成路线，并通过实验验证了其中的几条。计算机辅助多步反应规划的不断报道，包括算法开发和实验应用，凸显了该领域在合成化学家工作中的重要性日益增加。

(2) Reactivity prediction of single step transformation

(2) 单步转化的反应性预测

For molecular synthesis, although many reactions appear theoretically feasible, the reactions that can actually achieve uniformly high reactivity and selectivity are exceptionally rare [158-160]. More commonly, most reactions only achieve the desired efficiency and selectivity under a delicate combination of substrate, catalyst, and conditions. Therefore, accurate evaluation of the reactivity and selectivity of singlestep transformation is equally crucial for the successful design of molecular synthesis [161,162]. However, due to the vast molecular structural space and the multitude of controlling factors, there is no simple formulaic equation capable of quantitatively describing the universal laws of molecular synthesis. The QSAR of single-step transformation still remains one of the core challenges in AI synthesis [163-165]. Facing this challenge, traditional research has typically relied on experience-driven strategies: by summarizing the available data, synthetic chemists are able to derive a local structure-activity relationship for the specific target, which is then used for the rational design and improvement of synthetic transformation. However, this empirical approach lacks precision and predictive power, and conflicting rules can exist. This situation makes random selection and trial-and-error inevitable in designing and screening actual synthetic explorations.

对于分子合成，尽管许多反应在理论上看似可行，但实际上能够始终实现高反应性和选择性的反应却极为罕见[158 - 160]。更常见的情况是，大多数反应仅在底物、催化剂和条件的精细组合下才能达到所需的效率和选择性。因此，准确评估单步转化的反应性和选择性对于分子合成的成功设计同样至关重要[161,162]。然而，由于庞大的分子结构空间和众多的控制因素，不存在能够定量描述分子合成普遍规律的简单公式化方程。单步转化的定量构效关系仍然是AI合成中的核心挑战之一[163 - 165]。面对这一挑战，传统研究通常依赖经验驱动的策略:通过总结现有数据，合成化学家能够得出特定目标的局部结构 - 活性关系，然后将其用于合成转化的合理设计和改进。然而，这种经验方法缺乏精确性和预测能力，并且可能存在相互矛盾的规则。这种情况使得在设计和筛选实际合成探索中不可避免地需要进行随机选择和试错。

Recently, data-driven approaches have brought a new perspective to solve the problem of single-step QSAR prediction [127,166-168]. Benefiting from advanced AI algorithms and rich chemical data, a series of studies have shown that ML models are able to accurately predict reaction yields and selectivities [169-175], even surpassing the judgment of experienced chemists in some cases [12,176]. More importantly, these models can assist chemists in efficiently screening new catalysts for target reactions [177-179], providing powerful AI tools for molecular synthesis. These studies revealed the remarkable potential of ML technology in synthetic chemistry, promising to accelerate the process from the development of synthetic methods to the discovery of functional molecules.

最近，数据驱动的方法为解决单步定量构效关系预测问题带来了新的视角[127,166 - 168]。受益于先进的AI算法和丰富的化学数据，一系列研究表明机器学习模型能够准确预测反应产率和选择性[169 - 175]，在某些情况下甚至超过了经验丰富的化学家的判断[12,176]。更重要的是，这些模型可以帮助化学家有效地筛选目标反应的新催化剂[177 - 179]，为分子合成提供强大的AI工具。这些研究揭示了机器学习技术在合成化学中的巨大潜力，有望加速从合成方法开发到功能分子发现的过程。

Figure 4 Representative applications of SYNTHIA. (a) Highlighted complex organic molecules whose SYNTHIA’s predicted synthetic routes have been experimentally verified. (b) Calculation of the graph edit distance between two synthetic intermediates based on bond connections. Reproduced with permission from Ref. [45]. Copyright 2023, American Association for the Advancement of Science. (c) 3-step synthesis of (-)-stemoamide inspired by SYNTHIA (color online).

图4 SYNTHIA的代表性应用。(a) SYNTHIA预测的合成路线已通过实验验证的突出显示的复杂有机分子。(b) 基于键连接计算两个合成中间体之间的图编辑距离。经参考文献[45]许可转载。版权所有2023，美国科学促进会。(c) 受SYNTHIA启发的(-)-百部酰胺的三步合成(在线彩色图)。

For the palladium-catalyzed Buchwald-Hartwig cross-coupling reactions, Doyle et al. [180] demonstrated the potential of ML in predicting reaction yields. Utilizing a high-throughput synthetic platform, they reliably evaluated the yields of 4,140 reactions comprising a diverse range of substrates, catalysts, additives, and bases (Figure 5a). Through quantum chemistry computations and customized scripts, a series of atomic, molecular, and vibrational descriptors were automatically generated. Employing a random forest regression algorithm, they achieved an ${R}^{2}$ of 0.92 and a root mean square error (RMSE) of 7.8% across a 70% (training)/30% (validation) data split. Furthermore, the trained ML model is able to predict the outcomes for unseen additives, showcasing the extrapolative predictive power of the established yield model. Interestingly, the model interpretation revealed that the descriptors of isoxazoles were crucial for yield predictions. Following this mechanistic hint, the authors subsequently discovered that the active iso-xazoles were able to inhibit palladium’s catalytic activity through oxidative addition. This study, combining high-throughput experimentation with ML modeling, unveiled the attractive potential of data-driven research paradigm for reaction design and screening. It provides a powerful AI tool for evaluating the productivity of Buchwald-Hartwig cross-coupling reactions and enriches the understanding of the reaction mechanism.

对于钯催化的布赫瓦尔德-哈特维希交叉偶联反应，多伊尔等人[180]证明了机器学习在预测反应产率方面的潜力。他们利用高通量合成平台，可靠地评估了4140个反应的产率，这些反应包括各种底物、催化剂、添加剂和碱(图5a)。通过量子化学计算和定制脚本，自动生成了一系列原子、分子和振动描述符。采用随机森林回归算法，在70%(训练)/30%(验证)的数据划分中，他们实现了0.92的${R}^{2}$和7.8%的均方根误差(RMSE)。此外，训练好的机器学习模型能够预测未见过的添加剂的结果，展示了所建立的产率模型的外推预测能力。有趣的是，模型解释表明异恶唑的描述符对产率预测至关重要。遵循这一机理线索，作者随后发现活性异恶唑能够通过氧化加成抑制钯的催化活性。这项研究将高通量实验与机器学习建模相结合，揭示了数据驱动的研究范式在反应设计和筛选方面的诱人潜力。它为评估布赫瓦尔德-哈特维希交叉偶联反应的生产率提供了一个强大的人工智能工具，并丰富了对反应机理的理解。

By merging automation and ML modeling, Liao and colleagues [181] achieved selective Pd-catalyzed functionali-zation of sterically hindered aromatic meta-C-H bonds. They employed a synergistic protocol combining photoinduced C-H carboxylation, carboxy-directed Pd-catalyzed C-H func-tionalization, and microwave-assisted decarboxylation, using ${\mathrm{{CO}}}_{2}$ as a traceless director for targeted meta $\mathrm{C} - \mathrm{H}$ functionalization (Figure 5b). Through high-throughput experiments, they efficiently executed 1,032 reactions to explore a remarkable substrate scope, thereby providing comprehensive insights into the reaction’s synthetic potential. With this dataset in hand, they developed a yield prediction model using a message-passing neural network with pre-training from the USPTO dataset, which achieved an ${R}^{2}$ of 0.750 and an MAE of 7.2% in 5-fold cross-validation. In addition, this model is able to accurately predict the reaction outcome for unseen substrates, demonstrating significant advantages of high-throughput experimentation and ML-assisted yield prediction in exploring novel synthetic reactions.

通过将自动化与机器学习建模相结合，廖及其同事[181]实现了位阻芳香族间位C-H键的选择性钯催化官能团化。他们采用了一种协同方案，将光诱导C-H羧化、羧基导向的钯催化C-H官能团化和微波辅助脱羧相结合，使用${\mathrm{{CO}}}_{2}$作为无痕导向基团进行靶向间位$\mathrm{C} - \mathrm{H}$官能团化(图5b)。通过高通量实验，他们高效地进行了1032个反应，以探索广泛的底物范围，从而全面了解了该反应的合成潜力。有了这个数据集，他们使用来自美国专利商标局数据集进行预训练的消息传递神经网络开发了一个产率预测模型，该模型在5折交叉验证中实现了0.750的${R}^{2}$和7.2%的平均绝对误差(MAE)。此外，该模型能够准确预测未见过的底物的反应结果，证明了高通量实验和机器学习辅助产率预测在探索新型合成反应中的显著优势。

Figure 5 Selected ML yield prediction studies of organic transformation. (a) ML approach and performance of yield prediction of Pd-catalyzed Buchwald-Hartwig cross-coupling reactions. (b) ML approach and performance of yield prediction of Pd-catalyzed functionalization of sterically hindered aromatic meta-C-H bonds (color online).

图5有机转化的选定机器学习产率预测研究。(a)钯催化的布赫瓦尔德-哈特维希交叉偶联反应的机器学习方法和产率预测性能。(b)位阻芳香族间位C-H键的钯催化官能团化的机器学习方法和产率预测性能(在线彩色)。

Differing from the complete HTE dataset of chemical spaces, the synthetic exploration in real-world applications tends to be sparse and significantly more diverse in molecular selections. To investigate whether ML models could meet the challenge of predicting such scenarios, Wiest and colleagues [182] extracted and processed the data from As-traZeneca’s ELNs, creating a dataset for Buchwald-Hartwig reactions. This dataset included 781 reactions involving 340 aromatic halides, 260 amines, 24 ligands, 15 bases, and 15 solvents. Moreover, it contained a substantial number of low-yield or nonproductive reactions, with 39.9% yielding no product. This ELN dataset reflected the reality of synthetic transformation in pharmaceutical applications, presenting a significant challenge for ML modeling. The authors found that all attempted models, including the classic regression algorithms using RDKit features and more advanced Yield-BERT [183] model, failed to provide meaningful predictions, with the best model achieving an ${R}^{2}$ of only 0.266 . This finding, contrasting with the success of similar models on HTE datasets, indicated that ML models still need significant improvement in handling real-world synthetic scenarios and highlighted the need for caution in modeling with legacy yield data.

与化学空间的完整高通量实验(HTE)数据集不同，实际应用中的合成探索往往较为稀疏，分子选择也更加多样化。为了研究机器学习模型是否能够应对预测此类情况的挑战，维斯特及其同事[182]从阿斯利康的电子实验室笔记(ELN)中提取并处理了数据，创建了一个布赫瓦尔德-哈特维希反应的数据集。该数据集包括781个反应，涉及340种芳香卤化物、260种胺、24种配体、15种碱和15种溶剂。此外，它还包含大量低产或无产物的反应，39.9%的反应没有产物。这个ELN数据集反映了药物应用中合成转化的实际情况，对机器学习建模提出了重大挑战。作者发现，所有尝试的模型，包括使用RDKit特征的经典回归算法和更先进的Yield-BERT[183]模型，都未能提供有意义的预测，最好模型的${R}^{2}$仅为0.266。这一发现与类似模型在HTE数据集上的成功形成对比，表明机器学习模型在处理实际合成场景方面仍需大幅改进，并强调了在使用传统产率数据进行建模时需要谨慎。

Also targeting the challenge of biased distributions in literature data, Glorius et al. [184] noticed the importance of negative data in structure-activity relationship modeling. They found that, despite having up to 190,000 yield data from literature, models still struggled to achieve reliable predictions. Whether using traditional modeling methods or Yield-BERT [183] models, none could provide meaningful regression results, which is consistent with the above study from Wiest. By manipulating the data extraction and including additional random noise, the Glorius group attributed these poor modeling results primarily to the bias in literature data rather than the noise in experimental data. This bias stemmed from both selective sampling of reaction spaces in literature reports and a tendency to publish positive reaction outcomes. To address this issue, the authors proposed two strategies: purposefully conducting additional experiments to gather data on low-performing synthetic space, and data augmentation to help mitigate the issues caused by overly biased sampling. This work further revealed the importance of data distribution in synthetic chemistry modeling, highlighting the critical need for comprehensive, diversified, and fair evaluations and reporting in synthetic investigations.

同样针对文献数据中偏差分布的挑战，格洛里厄斯等人[184]注意到负数据在构效关系建模中的重要性。他们发现，尽管有来自文献的多达190,000个产率数据，但模型仍难以实现可靠的预测。无论是使用传统建模方法还是Yield - BERT[183]模型，都无法提供有意义的回归结果，这与维斯特的上述研究一致。通过操纵数据提取并引入额外的随机噪声，格洛里厄斯团队将这些糟糕的建模结果主要归因于文献数据中的偏差，而非实验数据中的噪声。这种偏差源于文献报道中反应空间的选择性采样以及发表正反应结果的倾向。为解决这个问题，作者提出了两种策略:有目的地进行额外实验以收集低性能合成空间的数据，以及数据增强以帮助减轻过度偏差采样导致的问题。这项工作进一步揭示了数据分布在合成化学建模中的重要性，强调了在合成研究中进行全面、多样化和公平评估及报告的迫切需求。

(3) Selectivity prediction of single-step transformation

(3) 单步转化的选择性预测

As a key component of structure-performance relationships, selectivity is also a crucial target for ML predictions in synthetic chemistry [161,164,166,185,186]. To realize the data-driven prediction of asymmetric catalysis, the Denmark group [187] reported the successful ML application in BI-NOL phosphoric acid (BPA)-catalyzed asymmetric addition of imines (Figure 6a), demonstrating the advantages of AI in solving stereoselectivity problems. The authors introduced innovative designs in both data selection and ML modeling. For data selection, they proposed a concept called universal training set (UTS), employing the Kennard-Stone algorithm to select chemically representative substances within a space composed of steric and electronic descriptors. This selection method, independent of reaction and mechanistic understanding, is solely based on the physicochemical properties of studied molecules, thereby rendering the chosen BPA set broadly applicable for modeling of BPA-involved transformations. In addition, to accurately characterize the complex steric environment of BPA molecules, a novel stereochemical descriptor called “average steric occupancy” (ASO) was developed. This descriptor is based on molecular occupancy at grid points within a cubic lattice: for each grid point, a value of 0 or 1 is assigned based on whether molecules occupy this position. The grid values were subsequently averaged across all conformers to generate the uniformed, high-dimensional ASO descriptor representing the steric environment of the molecule. Integrating these innovative designs, the authors attempted ML predictions of enantioselectivity on a dataset comprising 43 BPAs, 5 imines, and 5 thiols, totaling 1,075 reactions. The constructed neural network model precisely predicted the target en-antioselectivity, with a mean absolute deviation (MAD) of about 0.15 kcal/mol. Moreover, the model’s reliability was validated in several out-of-sample and out-of-range tasks, accurately predicting unseen catalysts and successfully differentiating superior ones. This work, with the innovative workflow and descriptor designs, provides a key reference for data-driven modeling of stereoselectivity, highlighting the potential of AI technology in addressing asymmetric challenges.

作为结构 - 性能关系的关键组成部分，选择性也是合成化学中机器学习预测的关键目标[161,164,166,185,186]。为实现不对称催化的数据驱动预测，丹麦团队[187]报道了在BI - NOL磷酸(BPA)催化的亚胺不对称加成反应中的成功机器学习应用(图6a)，展示了人工智能在解决立体选择性问题方面的优势。作者在数据选择和机器学习建模方面都引入了创新设计。对于数据选择，他们提出了一个名为通用训练集(UTS)的概念，采用肯纳德 - 斯通算法在由空间和电子描述符组成的空间内选择具有化学代表性的物质。这种选择方法独立于反应和机理理解，仅基于所研究分子的物理化学性质，从而使所选的BPA集广泛适用于涉及BPA的转化建模。此外，为了准确表征BPA分子复杂的空间环境，开发了一种名为“平均空间占有率”(ASO)的新型立体化学描述符。该描述符基于立方晶格内网格点处的分子占有率:对于每个网格点，根据分子是否占据该位置分配0或1的值。随后对所有构象异构体的网格值进行平均，以生成代表分子空间环境的统一的高维ASO描述符。综合这些创新设计，作者尝试对包含43种BPA、5种亚胺和5种硫醇的数据集(共1075个反应)进行对映选择性的机器学习预测。构建的神经网络模型精确预测了目标对映选择性，平均绝对偏差(MAD)约为0.15 kcal/mol。此外，该模型的可靠性在几个样本外和范围外任务中得到了验证，准确预测了未见过的催化剂并成功区分了优质催化剂。这项工作凭借创新的工作流程和描述符设计，为立体选择性的数据驱动建模提供了关键参考，突出了人工智能技术在应对不对称挑战方面的潜力。

Interestingly, Sigman and colleagues [188] also explored the stereoselectivity prediction of BPA-catalyzed imine addition reactions from the perspective of multivariate linear regression using physical organic parameters (Figure 6b). They posited that by uncovering the common mechanistic features of all reaction components, a comprehensive understanding of the factors controlling the reactivity and selectivity could be achieved. With this comprehensive set of physical organic controlling factors and leveraging statistical modeling, predictions for different structural motifs within a single model became feasible. For the studied BPA catalysis, the authors systemically parameterized the involved reaction components and catalysts from a physical organic chemistry standpoint, leading to the compilation of 313 parameters expressing steric and electronic effects. Based on this, multivariate linear regressions were made on an enantioselec-tivity dataset of 367 reactions compiled from relevant literature. The linear model achieved the remarkable performance of ${R}^{2}$ close to 0.9 . Notably, when applied to the exact same dataset published by Denmark [187], Sigman’s model also achieved excellent predictive performance, accurately identifying the selective catalysts. This highlights that the mechanism-based physical organic chemistry parameters can enable powerful QSAR prediction of focal datasets without the usage of sophisticated regression algorithms, which provides an alternative approach for quantitative predictions of stereoselectivity.

有趣的是，西格曼及其同事[188]也从使用物理有机参数的多元线性回归角度探索了BPA催化的亚胺加成反应的立体选择性预测(图6b)。他们认为，通过揭示所有反应组分的共同机理特征，可以全面理解控制反应性和选择性的因素。有了这一整套物理有机控制因素并利用统计建模，在单个模型中对不同结构基序进行预测就变得可行。对于所研究的BPA催化反应，作者从物理有机化学角度对涉及的反应组分和催化剂进行了系统参数化，得出了313个表达空间和电子效应的参数。基于此，对从相关文献汇编的367个反应的对映选择性数据集进行了多元线性回归。线性模型实现了${R}^{2}$接近0.9的显著性能。值得注意的是，当应用于丹麦团队[187]发表的完全相同的数据集时，西格曼的模型也取得了出色的预测性能，准确识别了选择性催化剂。这突出表明基于机理的物理有机化学参数无需使用复杂的回归算法就能对重点数据集进行强大的定量构效关系预测，这为立体选择性的定量预测提供了一种替代方法。

Figure 6 Selected ML enantioselectivity prediction studies of chiral phosphoric acid-catalyzed imine addition. (a) Data distribution, molecular descriptors with the design of ASO, and model performances. (b) Data distribution, physical organic descriptors, and the model performances using the multivariate linear regression approach (color online).

图6手性磷酸催化的亚胺加成反应的机器学习对映选择性预测研究选例。(a) 数据分布、带有ASO设计的分子描述符以及模型性能。(b) 数据分布、物理有机描述符以及使用多元线性回归方法的模型性能(在线彩色图)。

Exciting advances in stereoselectivity prediction were also made for transition metal catalysis [165,189,190]. Focusing on the asymmetric hydrogenation of olefins [189], Hong and colleagues [191] reported a productive ML modeling study. Due to the highly sparse and biased nature of the selected datasets from literature, a novel modeling strategy called “hierarchical learning” was proposed to overcome the bias and achieve extrapolative prediction. The core idea is to view the structure-selectivity relationship as a superposition of a universal relationship and local perturbations. A base model representing the universal structure-activity relationship is learned through the representative data samplings, followed by training a delta model with the neighboring data close to the target reaction, thus learning the perturbations of the relationship. The superposition of layered models yields the final prediction, which can be considered as an approach to transfer learning. This transfer learning strategy achieved excellent prediction in the asymmetric hydrogenation of olefins, requiring only limited reaction data of the target alkene substrate for satisfying modeling. The collaboration between Hong and Ackermann further applied the hierarchical learning strategy to explore the holistic synthetic space of electrochemical Pd-catalyzed C-H alkenylation (Figure 7a) [192], systematically studying the enantioselec-tivities of 846,720 reaction combinations, which demonstrated the appealing advantages of data-driven method in achieving the comprehensive knowledge of synthetic space. They also utilized this transfer learning protocol in virtual catalyst screening for asymmetric Co-catalyzed C-H alke-nylation, which successfully predicted and verified an intriguing chiral carboxylic acid with excellent enantioselectivity [193].

在过渡金属催化的立体选择性预测方面也取得了令人兴奋的进展[165,189,190]。以烯烃的不对称加氢反应为例[189]，Hong及其同事[191]开展了一项卓有成效的机器学习建模研究。鉴于从文献中选取的数据集具有高度稀疏和有偏的特性，他们提出了一种名为“分层学习”的新型建模策略，以克服偏差并实现外推预测。其核心思想是将结构-选择性关系视为普遍关系和局部扰动的叠加。通过代表性数据采样学习代表普遍结构-活性关系的基础模型，然后用接近目标反应的相邻数据训练增量模型，从而学习该关系的扰动。分层模型的叠加产生最终预测，这可被视为一种迁移学习方法。这种迁移学习策略在烯烃的不对称加氢反应中取得了出色的预测效果，仅需目标烯烃底物的有限反应数据即可满足建模需求。Hong和Ackermann的合作进一步将分层学习策略应用于探索电化学钯催化的C-H烯基化反应的整体合成空间(图7a)[192]，系统地研究了846,720种反应组合的对映选择性，展示了数据驱动方法在全面了解合成空间方面的诱人优势。他们还将这种迁移学习协议用于不对称钴催化的C-H烯基化反应的虚拟催化剂筛选，成功预测并验证了一种具有出色对映选择性的有趣手性羧酸[193]。

In addition to enantioselectivity, ML modeling has also been applied to other categories of stereoselectivity [172,194,195]. One representative study is Grzybowski’s work [194] in Diels-Alder reaction (Figure 7b). Using physical organic descriptors, the authors successfully predicted the major regio-, site-, and diastereoisomers using ML modeling. They found that using physical organic descriptors, as opposed to naive molecular fingerprints, significantly improved the model performance. By capturing electronic effects with Hammett constants and steric properties with TSEI indices, the chemical descriptors combined with a random forest classifier achieved an excellent prediction accuracy for regio- (93.6%), site- (91.3%), and dia-stereoselectivities (89.2%). The authors further demonstrated that the prediction performance of the Hammett-TSEI-based random forest classifier received less effect by the dataset partitioning compared to other encodings, indicating that the physical organic descriptors can enable the model to learn the organic structure-performance relationship and accurately predict outcomes for compounds unseen during model training.

除了对映选择性，机器学习建模还被应用于其他类型的立体选择性研究[172,194,195]。一项具有代表性的研究是Grzybowski在狄尔斯-阿尔德反应中的工作194。作者使用物理有机描述符，通过机器学习建模成功预测了主要的区域、位点和非对映异构体。他们发现，与简单的分子指纹相比，使用物理有机描述符能显著提高模型性能。通过用哈米特常数捕捉电子效应，用TSEI指数捕捉空间性质，这些化学描述符与随机森林分类器相结合，在区域选择性(93.6%)、位点选择性(91.3%)和非对映选择性(89.2%)方面取得了出色的预测准确率。作者进一步证明，与其他编码方式相比，基于哈米特-TSEI的随机森林分类器的预测性能受数据集划分的影响较小，这表明物理有机描述符能使模型学习有机结构-性能关系，并准确预测模型训练期间未见过的化合物的结果。

Regioselectivity prediction has also been realized using ML methods, as evidenced by a series of exciting advances in recent years [194,196-199]. The Hong group [197] has conducted ML studies on the regioselectivity of radical C-H functionalization of arenes (Figure 8a). Based on previous mechanistic understandings [200], they systematically computed the DFT barriers of the rate-determining step for a myriad of substrates, obtaining regioselectivity data for 9,438 reactions. They found that the physical organic descriptor, with only a few dozen dimensions, achieved satisfying regression results comparable to other typical higher-dimensional descriptors (smooth overlap of atomic positions (SOAP), atom-centered symmetry functions (ACSF), etc.). The trained random forest model accurately predicted the reaction sites with 94.2% accuracy and determined the degree of selectivity with 89.9% accuracy. Subsequently, the model’s predictions were compared with reported experimental results on complex polysubstituted aromatics. Despite being trained only on DFT data and not having been exposed to the experimental complex compounds, the model still performed with convincing accuracy. This work not only demonstrates that the DFT computation can provide a reliable data source for ML modeling of selectivity problems and can be directly applied to experimental prediction and verification, but emphasizes the importance of local physical organic descriptors of reaction sites in regioselectivity modeling.

近年来的一系列令人兴奋的进展证明，使用机器学习方法也实现了区域选择性预测[194,196 - 199]。Hong团队[197]对芳烃的自由基C-H官能化反应的区域选择性进行了机器学习研究(图8a)。基于之前的机理认识[200]，他们系统地计算了大量底物速率决定步骤的密度泛函理论(DFT)能垒，获得了9438个反应的区域选择性数据。他们发现，仅几十维的物理有机描述符取得了与其他典型的高维描述符(原子位置平滑重叠(SOAP)、以原子为中心的对称函数(ACSF)等)相当的令人满意的回归结果。训练后的随机森林模型以94.2%的准确率准确预测了反应位点，以89.9%的准确率确定了选择性程度。随后，将该模型的预测结果与关于复杂多取代芳烃的已报道实验结果进行了比较。尽管该模型仅在DFT数据上进行训练，且未接触过实验中的复杂化合物，但仍表现出令人信服的准确率。这项工作不仅表明DFT计算可为选择性问题的机器学习建模提供可靠的数据源，并可直接应用于实验预测和验证，还强调了反应位点的局部物理有机描述符在区域选择性建模中的重要性。

Jensen and colleagues [198] advanced the prediction modeling of regioselectivity for synthetic transformations, including aromatic C-H functionalization and C-X substitution. In their work, the implicit molecular representations derived from ML were combined with explicit quantum chemical properties (Figure 8b). The machine-learned molecular representation was realized by a GNN based on the Weisfeiler-Lehman network framework. Quantum chemical descriptors of organic molecules (atomic charges, Fukui indices, nuclear magnetic resonance (NMR) shielding constants, etc.) were calculated at the B3LYP/def2-SVP level using GFN2-xTB optimized structures. By combining these two types of molecular representations, they developed an ML model that accurately predicts the regioselectivity of the target reactions. To circumvent the time- and resource-consuming quantum chemical calculations, they demonstrated the appealing potential of combining multiple ML models. For this, they trained a directed message-passing neural network to predict QM properties of molecules, using these predictions instead of DFT-computed parameters as input for the selectivity model. This strategy provided an end-to-end model that can predict selectivity from SMILES within milliseconds. The evaluation showed that this fusion model achieved an accuracy of 89.7% for aromatic C-H functio-nalization reactions, 96.7% for aromatic C-X substitution reactions, and 97.2% for other substitution reactions. This work illustrates the complementary nature of data-driven representation and quantum chemical parameters for molecular encoding, while also highlighting the potential of on-the-fly quantum chemical property prediction and derived reactivity/selectivity modeling in synthetic chemistry.

詹森及其同事[198]推进了合成转化区域选择性的预测建模，包括芳香族C-H官能化和C-X取代反应。在他们的工作中，将机器学习得到的隐式分子表示与显式量子化学性质相结合(图8b)。基于魏斯费勒-莱曼网络框架的图神经网络实现了机器学习的分子表示。使用GFN2-xTB优化结构，在B3LYP/def2-SVP水平上计算有机分子的量子化学描述符(原子电荷、福井指数、核磁共振(NMR)屏蔽常数等)。通过结合这两种分子表示，他们开发了一个能准确预测目标反应区域选择性的机器学习模型。为了规避耗时且资源消耗大的量子化学计算，他们展示了组合多个机器学习模型的诱人潜力。为此，他们训练了一个定向消息传递神经网络来预测分子的量子力学性质，并使用这些预测结果而非密度泛函理论(DFT)计算的参数作为选择性模型的输入。这种策略提供了一个能在毫秒内从SMILES预测选择性的端到端模型。评估表明，该融合模型对芳香族C-H官能化反应的准确率为89.7%，对芳香族C-X取代反应的准确率为96.7%，对其他取代反应的准确率为97.2%。这项工作说明了数据驱动表示和量子化学参数在分子编码方面的互补性，同时也突出了在合成化学中即时量子化学性质预测以及由此得出的反应性/选择性建模的潜力。

Figure 7 Selected ML stereoselectivity prediction studies of Pd-catalyzed C-H alkenylation and Diels-Alder reaction. (a) Descriptor design, ML approach, and the model performances and predictions of the enantioselectivities of pallada-electrocatalyzed C-H activation. (b) Descriptor design, ML approach, and the model performances of the regio-, site-, and diastereoselectivities of Diels-Alder reaction (color online).

图7钯催化的C-H烯基化和狄尔斯-阿尔德反应的选定机器学习立体选择性预测研究。(a)描述符设计、机器学习方法以及钯电催化C-H活化对映选择性的模型性能和预测。(b)狄尔斯-阿尔德反应区域、位点和非对映选择性的描述符设计、机器学习方法以及模型性能(彩色在线)。

Hartwig and colleagues [201] successfully merged expert rules with data modeling to predict the regioselectivity of Ir-catalyzed C-H borylation reactions (Figure 8c). They combined literature results with specifically sampled low-selectivity data and representative intermolecular competition experiments, forming a data set for regioselectivity training. In selectivity modeling, they combined computational chemistry, data-driven approaches, and expert experience. Starting with a rough estimation of reaction activation barriers using the xTB method, they combined it with a partial least squares (PLS) model to predict the regioselectivity. Considering the limited scope of modeling data, they further implemented an expert rule to express the influence of neighboring substituents. Combining model predictions with rule-based corrections, they provided a predictive model for evaluating borylation sites. This approach maximized the benefits of computation, modeling, and expert experience, resulting in a highly effective selectivity prediction model. The model’s predictions aligned well with experimental verification across various compounds, which was also compared with predictions from experienced human chemists, outperforming them in tests on a few showcase complex compounds. This study demonstrates the complementary nature of human expertise and ML modeling, showing how expert rules can enhance and improve the predictive capabilities of ML models.

哈特维希及其同事[201]成功地将专家规则与数据建模相结合，以预测铱催化的C-H硼化反应的区域选择性(图8c)。他们将文献结果与专门采样的低选择性数据以及具有代表性的分子间竞争实验相结合，形成了一个用于区域选择性训练的数据集。在选择性建模中，他们结合了计算化学、数据驱动方法和专家经验。首先使用xTB方法对反应活化能垒进行粗略估计，然后将其与偏最小二乘(PLS)模型相结合来预测区域选择性。考虑到建模数据范围有限，他们进一步实施了一条专家规则来表达相邻取代基的影响。将模型预测与基于规则的校正相结合，他们提供了一个用于评估硼化位点的预测模型。这种方法最大限度地发挥了计算、建模和专家经验的优势，得到了一个高效的选择性预测模型。该模型的预测结果与各种化合物的实验验证结果高度吻合，并且还与经验丰富的化学家和预测结果进行了比较，在一些展示性复杂化合物的测试中表现优于他们。这项研究证明了人类专业知识与机器学习建模的互补性，展示了专家规则如何增强和提高机器学习模型的预测能力。

Figure 8 Selected ML regioselectivity prediction studies of organic transformation. (a) Descriptor design and ML performances of radical C-H func-tionalization of arenes. Reproduced with permission from Ref. [197]. Copyright 2020, John Wiley & Sons. (b) Regioselectivity prediction model of aromatic C-H functionalization and C-X substitution that combines the machine-learned representation by GNN and the calculated atomic descriptors. Reproduced with permission from Ref. [198]. Copyright 2021, Royal Society of Chemistry. (c) Regioselectivity prediction model of Ir-catalyzed C-H borylation that combines the xTB calculation, the ML regression, and the expert rule for neighboring substituent influence (color online).

图8有机转化的选定机器学习区域选择性预测研究。(a)芳烃自由基C-H官能化的描述符设计和机器学习性能。经参考文献[197]许可复制。版权所有2020，约翰威立父子公司。(b)结合图神经网络的机器学习表示和计算得到的原子描述符的芳香族C-H官能化和C-X取代反应区域选择性预测模型。经参考文献[198]许可复制。版权所有2021，皇家化学学会。(c)结合xTB计算、机器学习回归以及相邻取代基影响专家规则的铱催化C-H硼化反应区域选择性预测模型(彩色在线)。

(4) Reaction optimization

(4)反应优化

Data-driven reaction optimization allows an effective strategy to accelerate the discovery and improvement of synthetic processes by providing actionable recommendations for reaction conditions. This is an optimization problem in a defined chemical space, which does not fall into the category of regression or classification, necessitating an iterative workflow with experimental testing for correct and complete predictions in the face of chemical nuances. Traditionally, reaction optimization relies on chemists’ knowledge and design of experiment (DoE) methods, which, though automatable, do not harness the statistical value of the accumulated data and demand significant experiment efforts. In addition, the efficacy of these models relies heavily on the quality of the training data. Beker et al. [202] argued that noise and bias in literature data could hinder the creation of models that surpass literature popularity trends, emphasizing the importance of high-quality training data.

数据驱动的反应优化通过为反应条件提供可操作的建议，为加速合成过程的发现和改进提供了一种有效策略。这是一个在定义的化学空间中的优化问题，不属于回归或分类范畴，面对化学细微差别时需要通过实验测试进行迭代工作流程以获得正确和完整的预测。传统上，反应优化依赖于化学家的知识和实验设计(DoE)方法，这些方法虽然可以自动化，但没有利用积累数据的统计价值，并且需要大量的实验工作。此外，这些模型的有效性在很大程度上依赖于训练数据的质量。贝克尔等人[202]认为，文献数据中的噪声和偏差可能会阻碍创建超越文献流行趋势的模型，强调了高质量训练数据的重要性。

In 2018, the Aspuru-Guzik group [203] introduced Phoe-nics, an algorithm using Bayesian optimization for global optimization in chemical experimentation. Phoenics proposes new conditions based on previous observations, efficiently identifying optimal conditions and demonstrating applicability in complex case studies like the Oregonator, a nonlinear chemical reaction network. Sunoj and colleagues [204] developed an ML model for discovering catalysts in asymmetric hydrogenation, accurately predicting enantiomeric excess with an RMSE of ${8.4} \pm {1.8}$ using molecular parameters from 368 substrate-catalyst combinations. The model successfully predicted out-of-sample data, indicating potential for catalyst discovery and substrate selection. In 2021, Doyle, Adams, and colleagues [176] reported a Bayesian reaction optimization framework integrating algorithms into daily lab practices. They applied Bayesian optimization to Mitsunobu and deoxyfluorination reactions, enabling more efficient synthesis of value-added compounds through data-driven experimental decisions. Wang and colleagues [205] used ML to accelerate Cu catalyst discovery and optimization for ${\mathrm{{CO}}}_{2}$ reduction, identifying critical features and facilitating catalyst design and validation. In addition, the ML-assisted reaction optimization extends to materials. Norquist et al. [206] used support vector machine (SVM) for an ML-assisted materials discovery from failed experiments. Cooper’s group [207] integrated robotic experimentation and high-throughput computation to explore high-activity linear polymers for hydrogen evolution photocatalysts. It is also demonstrated that the ML model can recommend new reactions and generate hypotheses about crystal formation, emphasizing its versatility beyond traditional organic synthesis.

2018年，阿斯普鲁-古齐克团队[203]推出了Phoenics算法，该算法运用贝叶斯优化方法进行化学实验中的全局优化。Phoenics算法基于先前的观察结果提出新的条件，能高效识别最优条件，并在诸如俄勒冈振子(一种非线性化学反应网络)等复杂案例研究中展现出适用性。苏诺伊及其同事[204]开发了一种用于不对称氢化反应中发现催化剂的机器学习模型，利用368种底物-催化剂组合的分子参数，以${8.4} \pm {1.8}$的均方根误差准确预测对映体过量情况。该模型成功预测了样本外数据，显示出其在催化剂发现和底物选择方面的潜力。2021年，多伊尔、亚当斯及其同事[176]报告了一个将算法整合到日常实验室操作中的贝叶斯反应优化框架。他们将贝叶斯优化应用于光延反应和脱氧氟化反应，通过数据驱动的实验决策实现了更高效地合成增值化合物。王及其同事[205]利用机器学习加速用于${\mathrm{{CO}}}_{2}$还原的铜催化剂的发现和优化，识别关键特征并促进催化剂设计和验证。此外，机器学习辅助的反应优化还扩展到了材料领域。诺奎斯特等人[206]使用支持向量机(SVM)从失败的实验中进行机器学习辅助的材料发现。库珀团队[207]将机器人实验与高通量计算相结合，探索用于析氢光催化剂的高活性线性聚合物。研究还表明，机器学习模型可以推荐新的反应并生成关于晶体形成的假设，凸显了其超越传统有机合成的通用性。

4 AI applications in polymer synthesis

4 人工智能在聚合物合成中的应用

As key substances in the fields of molecular science and functional materials, the AI application in polymer synthesis has received wide interest and significant progress in recent years. Unlike small molecules with well-defined structures, polymers’ continuous molecular structure adds more possibilities in terms of structure and functionality. However, this also brings new challenges for data modeling. These challenges manifest both in how to represent the molecular structure of polymers in a data-driven format for ML models and in how to rationally establish quantitative relationships between the polymer synthesis process and the structure/ properties of the generated polymers. Furthermore, bioma-cromolecules like proteins and DNA are naturally programmable macromolecular machines. Addressing the synthesis of such biomacromolecules through data-driven solutions is not only a focal point of synthetic interest but also forms a key component in the field of bioinformatics. This section delves into the research on AI applications in polymer synthesis, with key directions highlighted in Figure 9. It aims to reveal the new opportunities AI brings to this field, as illustrated through discussions on representative works.

作为分子科学和功能材料领域的关键物质，近年来人工智能在聚合物合成中的应用受到了广泛关注并取得了重大进展。与结构明确的小分子不同，聚合物连续的分子结构在结构和功能方面增加了更多可能性。然而，这也给数据建模带来了新的挑战。这些挑战体现在如何以数据驱动的格式为机器学习模型表示聚合物的分子结构，以及如何合理地建立聚合物合成过程与生成的聚合物的结构/性能之间的定量关系。此外，蛋白质和DNA等生物大分子是天然的可编程大分子机器。通过数据驱动的解决方案来解决此类生物大分子的合成问题，不仅是合成领域的一个焦点，也是生物信息学领域的一个关键组成部分。本节深入探讨人工智能在聚合物合成中的应用研究，关键方向如图9所示。它旨在揭示人工智能给该领域带来的新机遇，通过对代表性作品的讨论进行说明。

4.1 Structure-property relationship prediction of polymer

4.1聚合物的结构-性能关系预测

Polymers constitute a significant class of materials that are ubiquitous, with applications spanning from daily products, including plastics and rubbers, to cutting-edge high-tech products in electronics, photonics, and biomedicines [208]. The highly tunable functionality of polymers arises from their remarkable diversity at both microscales (e.g., chemical composition, atomic-level connectivity) and macroscales (e.g., crystallinity, phase separation) [209]. Nevertheless, the vast and complex chemical and morphological spaces hinder the discovery of novel polymeric materials for specific purposes.

聚合物是一类重要的材料，广泛存在于日常生活用品(如塑料和橡胶)以及电子、光子和生物医药等前沿高科技产品中[208]。聚合物高度可调的功能源于其在微观尺度(如化学成分、原子级连接性)和宏观尺度(如结晶度、相分离)上的显著多样性[209]。然而，巨大而复杂的化学和形态空间阻碍了用于特定目的的新型聚合物材料的发现。

Materials science has witnessed transformative advancements through the integration of ML into polymer property prediction. ML techniques are able to leverage pre-existing experimental data and computational data from first-principles calculations or molecular dynamic simulations, establishing models for rapid and accurate predictions of the properties of new polymer materials [210]. The initial step in modeling the structure-property prediction of polymers involves defining their representation at atomic/molecular levels [211]. However, traditional text-based representations are labor-intensive, computationally demanding, and lack adaptability to diverse polymer classes. These drawbacks hinder the development of AI/ML pipelines for high-throughput applications. In parallel, it is also non-trivial to efficiently probe existing databases for further discoveries. Advanced ML frameworks provide promising prospects for bridging the gap by exploiting available data.

材料科学通过将机器学习集成到聚合物性能预测中取得了变革性进展。机器学习技术能够利用来自第一性原理计算或分子动力学模拟的现有实验数据和计算数据，建立模型以快速准确地预测新型聚合物材料的性能[210]。对聚合物结构-性能预测进行建模的第一步涉及在原子/分子水平上定义其表示[211]。然而，传统的基于文本的表示方法劳动强度大、计算要求高，并且缺乏对不同聚合物类别的适应性。这些缺点阻碍了用于高通量应用的人工智能/机器学习管道的发展。同时，有效地探索现有数据库以进行进一步发现也并非易事。先进的机器学习框架通过利用可用数据为弥合这一差距提供了有希望的前景。

As illustrated in Figure 10a, graph-based representations, which directly capture topological information of chemical structures, show favorability for property prediction tasks of polymers. Simine et al. [212] predicted ultraviolet-visible (UV-vis) spectroscopy of conjugated polymers directly from CG representations via a deep-learning model of long-short-term memory recurrent neural network. The approach demonstrated the potential to investigate organic optoelec-tronics through computational experiments invoking CG representations. Wang and colleagues [213] utilized CG representations to construct a high-dimensional design space. Bayesian optimization process efficiently explored this continuous space, offering comprehensive insights into molecular-level relationships influencing the lithium conductivity of polymer electrolytes. More recently, Aldeghi et al. [214] introduced a graph representation of molecular ensembles that captured key features including monomer compositions and chain architectures, using a weighted directed message-passing neural network tailored for polymer property prediction. The platform established a database of over 40,000 possible copolymers via calculation of electron affinity and ionization potential and achieved superior accuracy than off-the-shelf material informatics methods.

如图10a所示，基于图的表示直接捕获化学结构的拓扑信息，对聚合物的性能预测任务具有优势。Simine等人[212]通过长短期记忆循环神经网络的深度学习模型直接从CG表示预测共轭聚合物的紫外可见(UV-vis)光谱。该方法展示了通过调用CG表示的计算实验研究有机光电器件的潜力。Wang及其同事[213]利用CG表示构建了一个高维设计空间。贝叶斯优化过程有效地探索了这个连续空间，提供了对影响聚合物电解质锂电导率的分子水平关系的全面见解。最近，Aldeghi等人[214]引入了一种分子集合的图表示，该表示使用为聚合物性能预测量身定制的加权有向消息传递神经网络捕获包括单体组成和链结构在内的关键特征。该平台通过计算电子亲和能和电离势建立了一个包含超过40,000种可能共聚物的数据库，并实现了比现成的材料信息学方法更高的精度。

The Ramprasad group [209,215] constituted a user-friendly structure-property prediction platform named “Polymer Genome”. The informatics platform leveraged three hierarchical levels of fingerprints to capture features critical to describe a specific polymer property, which spanned from three-atom fragments, descriptors of the quantitative structure-property relationship type, to morphological descriptors such as the fraction of side-chain atoms (Figure 10b) [208]. Researchers utilized ML algorithms based on Gaussian process regression to generate prediction models that were implemented in the online platform “Polymer Genome”.

Ramprasad团队[209,215]构建了一个名为“聚合物基因组”的用户友好型结构-性能预测平台。该信息学平台利用三个层次的指纹来捕获描述特定聚合物性能至关重要的特征，这些特征从三原子片段、定量结构-性能关系类型的描述符到侧链原子分数等形态描述符(图10b)[208]。研究人员利用基于高斯过程回归的机器学习算法生成预测模型，并在在线平台“聚合物基因组”中实现。

Fingerprints as inputs for predictive models tend to attain the relatively best performance [216]. However, hierarchical handcrafted fingerprints, necessitating chemical intuition, consume an amount of time due to complicated computations for model training and inference. Recent advancements in NLP have established Transformer as a powerful AI framework for language modeling. SMILES strings, considered the “chemical language” of polymers, entitle Transformer-based models to the opportunity for application in polymer science. Xu and colleagues [217] introduced TransPolymer, a Transformer-based model benefitting from pretraining on a large unlabeled dataset. This model showcased the importance of chemical awareness in modeling polymer sequences, affording a robust tool for structural-property relationship exploration. Similarly, Kuenneth et al. [218] trained polyBERT on 100 million polymer SMILES strings of hypothetical polymers to function as a chemical linguist. Integrated into multitask deep neural networks, the fully machine-learned polyBERT fingerprints predicted polymer properties at unparalleled speed with unimpaired accuracy, surpassing the SOTA record of handcrafted Polymer Genome fingerprints (Figure 10c).

指纹作为预测模型的输入往往能获得相对最佳的性能[216]。然而，分层手工制作的指纹需要化学直觉，由于模型训练和推理的复杂计算而耗费大量时间。自然语言处理的最新进展使Transformer成为一种强大的语言建模人工智能框架。SMILES字符串被认为是聚合物的“化学语言”，使基于Transformer的模型有机会应用于聚合物科学。Xu及其同事[217]引入了TransPolymer，这是一种基于Transformer的模型，受益于在大型未标记数据集上的预训练。该模型展示了化学意识在聚合物序列建模中的重要性，为结构-性能关系探索提供了一个强大的工具。同样，Kuenneth等人[218]在1亿个假设聚合物的聚合物SMILES字符串上训练了polyBERT，使其作为化学语言学家发挥作用。集成到多任务深度神经网络中，完全机器学习的polyBERT指纹以无与伦比的速度预测聚合物性能，且准确性不受影响，超过了手工制作的聚合物基因组指纹的SOTA记录(图10c)。

Figure 9 Key directions of AI applications in polymer synthesis (color online).

图9聚合物合成中人工智能应用的关键方向(在线彩色图)。

Figure 10 Illustration of (a) graph-based representations and (b) Transformer-based models favorable for structure-property relationship prediction of polymers (color online).

图10(a)基于图的表示和(b)有利于聚合物结构-性能关系预测的基于Transformer的模型的示意图(在线彩色图)。

Deep-learning architectures have revolutionized structure-property prediction of polymers by automatically learning expressive representations (Figure 11a). Rahman et al. [219] proposed a CNN-based framework that predicted the critical mechanical property, namely pullout force, of carbon nano-tube-polymer interfaces. Park et al. [220] utilized GCNs for predicting the thermal and mechanical properties of polymers. They found that GCNs, especially when combined with neural network regression, could slightly outperform the widely used extended-connectivity circular fingerprint (ECFP) representation.

深度学习架构通过自动学习有表现力的表示彻底改变了聚合物的结构-性能预测(图11a)。Rahman等人[219]提出了一个基于卷积神经网络的框架，用于预测碳纳米管-聚合物界面的关键力学性能，即拔出力。Park等人[220]利用图卷积网络预测聚合物的热性能和力学性能。他们发现图卷积网络，特别是与神经网络回归相结合时，能够略微优于广泛使用的扩展连接圆形指纹(ECFP)表示。

Figure 11 Schematic representation of (a) deep-learning architectures and (b) transfer learning-based frameworks for polymer property prediction (color online).

图11(a)用于聚合物性能预测的深度学习架构和(b)基于迁移学习的框架的示意图(在线彩色图)。

In addition to the structure-property relationship prediction, correlations between chemical, electronic, mechanical, and thermodynamic properties offer alternative avenues for effective property prediction models. Transfer learning leveraging models based on interrelated properties proves promising for predicting target properties even with minimal data (Figure 11b). The Yoshida group [221] developed Xe-nonPy.MDL, a pretrained model library with over 140,000 models for diverse properties of organic small molecules, polymers, and inorganic materials. Their frameworks, exemplified by neural network models, demonstrated efficient property prediction for extremely small datasets. Through a multi-fidelity fusion strategy that addresses the limitation of experimental data in quantity and diversity, the Ramprasad group [222] trained the model on the low-fidelity but abundant data set employing group contribution methods to predict polymer crystallinity under high-fidelity accuracy. Later, the same group [223] advocated multi-task learning by exploiting intercorrelations between various property datasets, yielding efficient, scalable, and interpretable models for polymer property prediction.

除了结构-性能关系预测之外，化学、电子、机械和热力学性质之间的相关性为有效的性能预测模型提供了替代途径。利用基于相关性质的模型进行迁移学习，即使在数据极少的情况下，也有望预测目标性质(图11b)。吉田小组[221]开发了Xe-nonPy.MDL，这是一个预训练模型库，包含超过140,000个用于有机小分子、聚合物和无机材料各种性质的模型。以神经网络模型为例，他们的框架展示了对极小型数据集的高效性能预测。通过解决实验数据在数量和多样性方面的局限性的多保真度融合策略，兰普拉德小组[222]在低保真但丰富的数据集上使用基团贡献方法训练模型，以在高保真精度下预测聚合物结晶度。后来，同一小组[223]倡导通过利用各种性质数据集之间的相互关联进行多任务学习，从而产生用于聚合物性质预测的高效、可扩展且可解释的模型。

4.2 Target-orientated design of polymer

4.2 聚合物的目标导向设计

In addition to the forward structure-property relationship prediction, it would be ideal to inverse design the desired polymer with target property. This concept of inverse design presents a novel paradigm, departing from the traditional Edisonian method, which accompanies time- and labor-intensive exploration reliant on human intuition with inherent biases and knowledge limitations. This approach enables generating polymers with superior functionality or properties by navigating the chemical space informed by data-driven strategies. Two avenues, high-throughput screening and advanced ML algorithms, are recognized as pivotal protocols to achieve the target-oriented polymer design (Figure 12).

除了正向的结构-性能关系预测之外，理想的情况是对具有目标性能的所需聚合物进行逆向设计。这种逆向设计概念提出了一种新颖的范式，与传统的爱迪生式方法不同，传统方法依赖于人类直觉进行耗时费力的探索，存在固有的偏差和知识局限性。这种方法能够通过在数据驱动策略所提供的化学空间中进行探索，生成具有卓越功能或性能的聚合物。高通量筛选和先进的机器学习算法这两条途径，被认为是实现目标导向聚合物设计的关键方案(图12)。

In the context of high-throughput screening, researchers should narrow the chemical space by defining the inputs of polymer fragments and adjoining rules based on their prior knowledge and chemical intuition, which would simultaneously ensure the validation of combining building blocks. For example, the Ramprasad group leveraged a polymer database derived from first principles, exploring the linear combination of 7 basic building blocks to recommend novel dielectric polymers [224]. In a similar manner, Afzal et al. [225] identified polyimides with exceptional refractive index values via high-throughput virtual screening, which could access a massive library of polyimide structures composed of 29 building blocks. However, integrating polymer fragments as inputs is prone to neglecting interaction between polymer chains and other influential factors in realistic production.

在高通量筛选的背景下，研究人员应根据其先验知识和化学直觉，通过定义聚合物片段的输入和邻接规则来缩小化学空间，这将同时确保组合构建块的有效性。例如，Ramprasad团队利用基于第一性原理的聚合物数据库，探索7种基本构建块的线性组合，以推荐新型介电聚合物[224]。以类似方式，Afzal等人[225]通过高通量虚拟筛选确定了具有优异折射率值的聚酰亚胺，该筛选可以访问由29种构建块组成的大量聚酰亚胺结构库。然而，将聚合物片段作为输入容易忽略聚合物链之间的相互作用以及实际生产中的其他影响因素。

Active learning and the derived AI-driven space exploration have also been utilized in the search for polymer candidates. Bayesian optimization, a noise-tolerant and global optimization strategy free from assumptions of functional forms, has also been implemented in polymer design. The workflow utilized by Wu et al. [226] overcame the challenge of limited data by incorporating transfer learning coupled with BO process. The approach empowered attaining quantitative structure-property relationships for thermal conductivity, which provided candidates possessing comparable thermal conductivities to those of SOTA non-composite thermo-plastics. Kim et al. [227] employed the genetic algorithm process that mimics the natural selection, creating over 100 novel polymers with a high glass transition temperature $\left( {T}_{\mathrm{g}}\right) > {500}\mathrm{\;K}$ and bandgap $\left( {E}_{\mathrm{g}}\right) > 6\mathrm{{eV}}$ , which are suitable for dielectric materials for high-temperature capacitors. Moreover, researchers suggested that optimized GA parameters and the biased initial population with prior knowledge could significantly improve the GA scheme. Zhou et al. [228] demonstrated that a non-periodic and nonintuitive sequence of PE-PP copolymers, which was generated through the genetic algorithm, outperformed regular block copolymers in thermal conductivity. Atomistic molecular dynamics then performed the fitness evaluation of each candidate by measuring its thermal conductivity.

主动学习和由此衍生的人工智能驱动的空间探索也已用于寻找聚合物候选物。贝叶斯优化是一种无需函数形式假设的抗噪声全局优化策略，也已在聚合物设计中得到应用。Wu等人[226]采用的工作流程通过结合迁移学习与贝叶斯优化过程，克服了数据有限的挑战。该方法能够实现热导率的定量结构-性能关系，从而提供了与最先进的非复合热塑性塑料具有可比热导率的候选物。Kim等人[227]采用了模拟自然选择的遗传算法过程，创建了100多种具有高玻璃化转变温度$\left( {T}_{\mathrm{g}}\right) > {500}\mathrm{\;K}$和带隙$\left( {E}_{\mathrm{g}}\right) > 6\mathrm{{eV}}$的新型聚合物，这些聚合物适用于高温电容器的介电材料。此外，研究人员表明，优化的遗传算法参数和具有先验知识的有偏初始种群可以显著改进遗传算法方案。Zhou等人[228]证明，通过遗传算法生成的非周期性和非直观的PE-PP共聚物序列在热导率方面优于常规嵌段共聚物。然后，原子分子动力学通过测量每个候选物的热导率对其进行适应性评估。

Figure 12 Schematic representation of high-throughput screening and advanced ML algorithms for target-oriented polymers (color online).

图12 用于目标导向聚合物的高通量筛选和先进机器学习算法的示意图(在线彩色)。

4.3 Design and optimization of polymer synthesis

4.3 聚合物合成的设计与优化

The advent of synthetic plastics in the last century revolutionized the chemical industry and the world at large. Polymer materials are now ubiquitous in our daily lives. However, traditional polymer synthesis is a process fraught with trial and error. Polymerization condition optimizing and catalyst screening are time- and resource-consuming endeavors. Moreover, this trial-and-error approach generates a significant amount of chemical waste, posing environmental concerns.

上世纪合成塑料的出现彻底改变了化学工业乃至整个世界。聚合物材料如今在我们的日常生活中无处不在。然而，传统的聚合物合成是一个充满试错的过程。聚合条件优化和催化剂筛选既耗时又耗资源。此外，这种试错方法会产生大量化学废物，引发环境问题。

Optimizing polymerization conditions is not a simple task with a singular focus. Chemists often need to balance parameters like chemical composition, molecular weight (MW), and dispersity $\left( D\right)$ for superior material properties. This multi-parameter, multi-objective optimization is a monumental task, compounded by the complexity of high-dimensional data, which often hinders chemists from precisely attaining diverse polymer targets. From this perspective, the potential for implementing AI and ML in polymer synthesis is enormous. By employing advanced data analysis techniques and predictive models, AI can assist scientists in rapidly identifying optimal polymerization conditions and catalysts, thereby reducing the cycle of experiments and saving time significantly.

优化聚合条件并非一项简单的单一任务。化学家通常需要平衡诸如化学成分、分子量(MW)和分散度$\left( D\right)$等参数，以获得优异的材料性能。这种多参数、多目标的优化是一项艰巨的任务，再加上高维数据的复杂性，常常阻碍化学家精确实现各种聚合物目标。从这个角度来看，在聚合物合成中应用人工智能和机器学习的潜力巨大。通过采用先进的数据分析技术和预测模型，人工智能可以帮助科学家快速确定最佳聚合条件和催化剂，从而减少实验周期并大幅节省时间。

However, the application of AI in polymer synthesis has been slower compared with its usage in optimizing small molecule organic synthesis. A primary reason is the lack of sufficient high-quality data in polymer research. The scarcity of data stems from the complexity of the polymerization process: The multi-parametric conditions and intricate polymerization mechanisms make computational simulations challenging, rendering simulated data unavailable. Additionally, stringent and sensitive polymerization conditions lead to significant variability in outcomes between different batches. These issues of data limitation pose challenges for ML modeling of polymer synthesis. Therefore, one of the key tasks for data-driven modeling of polymer synthesis relies on the acquisition of large quantities of high-quality, repeatable, and interpretable data that adhere to the FAIR (Findable, Accessible, Interoperable, and Reusable) principles [229].

然而，与在优化小分子有机合成中的应用相比，人工智能在聚合物合成中的应用进展较慢。一个主要原因是聚合物研究中缺乏足够的高质量数据。数据稀缺源于聚合过程的复杂性:多参数条件和复杂的聚合机制使得计算模拟具有挑战性，导致无法获得模拟数据。此外，严格且敏感的聚合条件导致不同批次之间的结果存在显著差异。这些数据限制问题给聚合物合成的机器学习建模带来了挑战。因此，聚合物合成数据驱动建模的关键任务之一依赖于获取大量符合FAIR(可查找、可访问、可互操作和可重用)原则的高质量、可重复且可解释的数据[229]。

One way to obtain data is searching from handbooks or literature. However, inconsistent and sometimes even contradictory results across different publications are not uncommon. High-throughput computational simulations or virtual screening is another approach. However, this approach highly depends on the computational power and is still challenging to predict experimental outcomes in a quantitative manner, especially for polymer synthesis. High-throughput experiment is a viable approach ensuring experimental consistency but relies on automation and is not suitable for experiments with long measurement times or complex material handling steps.

获取数据的一种方法是从手册或文献中搜索。然而，不同出版物之间结果不一致甚至相互矛盾的情况并不罕见。高通量计算模拟或虚拟筛选是另一种方法。然而，这种方法高度依赖计算能力，并且仍然难以定量预测实验结果，尤其是对于聚合物合成。高通量实验是确保实验一致性的可行方法，但依赖自动化，不适用于测量时间长或材料处理步骤复杂的实验。

Flow chemistry is currently the primary choice in AI-assisted polymerization processes optimization for its ability of real-time monitoring, time-dependent data acquisition on polymer MW and monomer conversion. For example, Junkers and colleagues developed an automated flow synthesis platform for polymer synthesis, coupled with real-time monitoring using gel permeation chromatography (GPC) [230], NMR [231], and Fourier-transform infrared spectroscopy [232]. This platform enables rapid and efficient screening of reaction parameters, including residence time, monomer concentration, polymerization degree, reaction temperature, and monomer conversion rate. They used single-objective ML optimization algorithms to dynamically adjust reaction parameters, optimizing the reaction to precisely control MW or monomer conversion rate, leading to significantly reduced experimental cycles and development time. To achieve multi-objective closed-loop optimization of polymer synthesis, Warren and colleagues demonstrated an ML-assisted automated flow polymerization synthesis platform that can autonomously determine optimal polymerization reaction conditions toward predetermined polymer properties [233]. It features a computer-controlled flow reactor that autonomously polymerizes, using real-time NMR and GPC for polymer characterization. This platform utilizes the Thompson Sampling Efficient Multi-Objective Optimization (TSEMO) algorithm to optimize reversible addition-fragmentation chain transfer (RAFT) polymerization of different monomers, exploring the trade-off between $D$ and monomer conversion rate. Hartman and colleagues [234] also combined automated microfluidics with ML to explore the reaction space of olefin free radical polymerization catalysts, accelerating the discovery of optimal catalytic efficiency conditions.

流动化学因其能够实时监测、获取聚合物分子量和单体转化率的时间相关数据，目前是人工智能辅助聚合过程优化的主要选择。例如，容克斯及其同事开发了一个用于聚合物合成的自动化流动合成平台，并结合使用凝胶渗透色谱(GPC)[230]、核磁共振(NMR)[231]和傅里叶变换红外光谱[232]进行实时监测。该平台能够快速有效地筛选反应参数，包括停留时间、单体浓度、聚合度、反应温度和单体转化率。他们使用单目标机器学习优化算法动态调整反应参数，优化反应以精确控制分子量或单体转化率，从而显著减少实验周期和开发时间。为了实现聚合物合成的多目标闭环优化，沃伦及其同事展示了一个机器学习辅助的自动化流动聚合合成平台，该平台可以朝着预定的聚合物性能自主确定最佳聚合反应条件[233]。它具有一个计算机控制的流动反应器，能够自主聚合，并使用实时核磁共振和凝胶渗透色谱对聚合物进行表征。该平台利用汤普森采样高效多目标优化(TSEMO)算法优化不同单体的可逆加成-断裂链转移(RAFT)聚合反应，探索$D$和单体转化率之间的权衡。哈特曼及其同事[234]还将自动化微流控技术与机器学习相结合，探索烯烃自由基聚合催化剂的反应空间，加速发现最佳催化效率条件。

In analyzing the relationship between polymerization parameters and outcomes, Chen and colleagues [235] developed an ML-assisted systematic polymerization planning (SPP) platform for intelligent control of polymer MW and $D$ . They constructed an ML model to analyze and optimize the reversible deactivation free radical polymerization process, combining multivariate analysis to uncover complex interactions between polymerization conditions for optimal polymerization condition prediction (Figure 13). Wilson and colleagues [236] also employed active learning and Bayesian optimization algorithms to accelerate the optimization of electrochemical atom transfer free radical polymerization reactions, which significantly improved the experiment efficiency.

在分析聚合参数与结果之间的关系时，陈及其同事[235]开发了一个机器学习辅助的系统聚合规划(SPP)平台，用于智能控制聚合物分子量和$D$。他们构建了一个机器学习模型来分析和优化可逆失活自由基聚合过程，结合多变量分析揭示聚合条件之间的复杂相互作用，以预测最佳聚合条件(图13)。威尔逊及其同事[236]还采用主动学习和贝叶斯优化算法来加速电化学原子转移自由基聚合反应的优化，这显著提高了实验效率。

4.4 End-to-end prediction of polymerization

4.4 聚合反应的端到端预测

Optimization of polymer conditions or catalysts for desired MW and $D$ , however, is often not the end of polymer synthesis in real-world practices. The ultimate goal of AI-assisted polymer synthesis, as previously mentioned, is to identify targeted (multi-)functions from high dimensional, enormous chemical spaces, elucidate the hidden structure-function relationships, and accelerate material design. Thus, many laboratories in recent years have been dedicated to the development of closed-loop high-throughput “design-synthesis-test-learn” toward end-to-end AI-assisted prediction from polymer synthesis to properties/functions. This approach, again, highly relies on robust high-throughput polymer synthesis platform amenable to automation and data digitalization for successive AI/ML.

然而，在实际应用中，为了获得所需的分子量和$D$而对聚合条件或催化剂进行优化，往往并非聚合物合成的终点。如前所述，人工智能辅助聚合物合成的最终目标是从高维、庞大的化学空间中识别目标(多)功能，阐明隐藏的结构-功能关系，并加速材料设计。因此，近年来许多实验室致力于开发闭环高通量“设计-合成-测试-学习”方法，以实现从聚合物合成到性能/功能的端到端人工智能辅助预测。这种方法同样高度依赖于强大的高通量聚合物合成平台，该平台需适用于自动化和数据数字化，以便后续进行人工智能/机器学习。

There are two primary strategies for high-throughput polymer synthesis, namely parallel copolymerization of various monomers and post-polymerization modifications. For the parallel copolymerization approach, the main challenges include: (1) polymerization reactions that are sensitive to moisture and/or air, increasing difficulties for automated liquid handling systems; (2) poor polymerization control, resulting in low repeatability and predictability; (3) limited flexibility in structural units, restricting monomer types and chemical space; (4) complex and inefficient posttreatment operations, difficult to pursue high throughput. To tackle these challenges, researchers have developed various water- and oxygen-resistant controlled free radical polymerization platforms, such as Enz-RAFT [237-240], oxygen-tolerant atom transfer radical polymerization (ATRP) [241], PET (photoinduced electron/energy transfer)-RAFT [242,243], and others [244-248], most of which overcome these issues and enable controlled high-throughput preparation of polymers. The second strategy is based on post-polymerization modification, which has been one of the primary methods for preparing high-throughput polymer libraries in recent years [249]. Efficient Huisgen cycloaddition [250], activated ester-amine coupling, thiol-ene reactions [251], and Michael addition are the most commonly used post-polymerization modification methods. The challenge with this strategy lies in the typically poor water solubility, instability, and difficulty in long-term storage of the precursor polymers.

高通量聚合物合成主要有两种策略，即各种单体的平行共聚和聚合后修饰。对于平行共聚方法，主要挑战包括:(1)聚合反应对水分和/或空气敏感，增加了自动化液体处理系统的难度；(2)聚合控制不佳，导致重复性和可预测性低；(3)结构单元的灵活性有限，限制了单体类型和化学空间；(4)复杂且低效的后处理操作，难以实现高通量。为应对这些挑战，研究人员开发了各种耐水和耐氧的可控自由基聚合平台，如酶催化可逆加成-断裂链转移聚合(Enz-RAFT)[237-240]、耐氧原子转移自由基聚合(ATRP)[241]、光诱导电子/能量转移(PET)-可逆加成-断裂链转移聚合(RAFT)[242,243]以及其他方法[244-248]，其中大多数克服了这些问题，能够实现聚合物的可控高通量制备。第二种策略基于聚合后修饰，这是近年来制备高通量聚合物库的主要方法之一[249]。高效的Huisgen环加成反应[250]、活化酯-胺偶联反应、硫醇-烯反应[251]以及迈克尔加成反应是最常用的聚合后修饰方法。该策略的挑战在于前体聚合物通常水溶性差、不稳定且难以长期储存。

Figure 13 An ML-assisted systematical polymerization planning (SPP) platform for polymer inverse design. Reproduced with permission from Ref. [235]. Copyright 2021, Science China Press (color online).

图13用于聚合物逆向设计的机器学习辅助系统聚合规划(SPP)平台。经参考文献[235]许可复制。版权所有2021，中国科学出版社(在线彩色版)。

For efficient discovery of polymers with certain functions, quick and convenient polymer purification methods are also needed in addition to high-throughput synthetic techniques. Gormley’s group [252] developed a gel filtration chromatography technique that rapidly and high-throughput purifies polymers, with over 95% removal of small molecule impurities and about 85% retention rate for 32 types of polymers. However, many polymer purification strategies (such as precipitation, extraction, and chromatographic separation) often depend on specific properties of the target polymers [253,254]. And due to the complexity of these processes, the purification step was sometimes skipped in some cases.

为了高效发现具有特定功能的聚合物，除了高通量合成技术外，还需要快速便捷的聚合物纯化方法。戈姆利团队[252]开发了一种凝胶过滤色谱技术，可快速、高通量地纯化聚合物，能去除超过95%的小分子杂质，对32种聚合物的保留率约为85%。然而，许多聚合物纯化策略(如沉淀、萃取和色谱分离)通常依赖于目标聚合物的特定性质[253,254]。由于这些过程的复杂性，在某些情况下有时会跳过纯化步骤。

Once a vast amount of structural and informational data were successfully generated through the high-throughput synthesis and characterization methods, the next key issue is how to effectively mine these data to guide new material design. Applying ML to identify key features from past data can guide studies on material structure-function relationships [255,256]. For instance, Reineke and colleagues [43,257] combined polymer design with parallel experimental work-flows to discover efficient polymers for intracellular ribonucleoprotein (RNP) delivery. Utilizing interpretable ML, they computed SHAP (Shapley additive explanations) for nine polymorphic features, uncovering the structure-function relationship behind editing efficiency, cytotoxicity, and RNP uptake, providing guidelines for designing polymer libraries based on RNP delivery. Bao et al. [258] reported an ML-assisted method that guides the design of full-color tunable emission trans-space charge transfer through-space charge transfer (TSCT) polymers. They synthesized 71 different chain length and type styrene polymers through ATRP, building Maximum Likelihood Expectation Multivariate Linear Regression (MLREM) and Bayesian Regularized Artificial Neural Network (BRANNLP) models to predict the photophysical properties of unknown TSCT polymers, exploring the relationship between structure and function. Olsen et al. [259] used high-throughput synthesis techniques to create a large library of 642 polyesters and polycarbonates, while developing a high-throughput clean area biodegradation test to assess the biodegradability of the polymers. They used ML models to interpret the structure-property relationships of polymer biodegradability. Knight and colleagues [260] designed and synthesized a series of polymers containing novel triphenylphosphine acrylamide monomers, using ML regression models to study the relationship between polymer properties and polymerization catalysis rates.

一旦通过高通量合成和表征方法成功生成了大量的结构和信息数据，下一个关键问题是如何有效地挖掘这些数据以指导新材料设计。应用机器学习从过去的数据中识别关键特征，可以指导对材料结构-功能关系的研究[255,256]。例如，赖内克及其同事[43,257]将聚合物设计与平行实验工作流程相结合，以发现用于细胞内核糖核蛋白(RNP)递送的高效聚合物。他们利用可解释的机器学习方法，计算了九种多态特征的SHAP(Shapley加法解释)值，揭示了编辑效率、细胞毒性和RNP摄取背后的结构-功能关系，为基于RNP递送的聚合物库设计提供了指导。鲍等人[258]报道了一种机器学习辅助方法，用于指导全色可调发射跨空间电荷转移(TSCT)聚合物的设计。他们通过原子转移自由基聚合(ATRP)合成了71种不同链长和类型的苯乙烯聚合物，建立了最大似然期望多元线性回归(MLREM)和贝叶斯正则化人工神经网络(BRANNLP)模型，以预测未知TSCT聚合物的光物理性质，探索结构与功能之间的关系。奥尔森等人[259]使用高通量合成技术创建了一个包含642种聚酯和聚碳酸酯的大型库，同时开发了一种高通量清洁区域生物降解测试来评估聚合物的生物降解性。他们使用机器学习模型来解释聚合物生物降解性的结构-性质关系。奈特及其同事[260]设计并合成了一系列含有新型三苯基膦丙烯酰胺单体 的聚合物，使用机器学习回归模型研究聚合物性质与聚合催化速率之间的关系。

In recent years, the ML-assisted closed-loop HTE has shown immense potential in the discovery of new materials [261-263] (Figure 14a). For example, Leibfarth and colleagues [264] combined ML with flow polymerization, enhancing the magnetic resonance signal strength of fluorinated polymers through only about 300 experiments and discovering previously unreported structure-effect relationships (Figure 14b). Gormley et al. [265], through a closed-loop high-throughput polymerization and active learning strategy, rapidly discovered polymers for neuroregeneration research that could protect proteins, as well as designed stable proteinase-active random copolymers [266] (Figure 14c). Lu and colleagues [267] established a high-throughput post-polymerization modification platform for selenium-containing polypeptides synthesis. By incorporating ML algorithms, they were able to efficiently explore the functional chemical space of 600 random copolymers for desired functions such as enzyme-like catalysis without much prior knowledge in four days (Figure 15). It is foreseeable that the deep integration of high-throughput technology and ML will have a significant impact on polymers and will aid in accelerating the discovery of materials in key areas.

近年来，机器学习辅助的闭环高通量实验在新材料发现方面展现出巨大潜力261 - 263。例如，Leibfarth及其同事[264]将机器学习与流动聚合相结合，仅通过约300次实验就增强了氟化聚合物的磁共振信号强度，并发现了此前未报道的结构 - 效应关系(图14b)。Gormley等人[265]通过闭环高通量聚合和主动学习策略，迅速发现了可用于神经再生研究的能保护蛋白质的聚合物，还设计了稳定的蛋白酶活性无规共聚物266。Lu及其同事[267]建立了用于含硒多肽合成的高通量聚合后修饰平台。通过纳入机器学习算法，他们能够在四天内高效探索600种无规共聚物的功能化学空间，以寻找如类酶催化等所需功能，且无需太多先验知识(图15)。可以预见，高通量技术与机器学习的深度整合将对聚合物产生重大影响，并有助于加速关键领域材料的发现。

Figure 14 (a) An ML-assisted design-build-test-learn closed-loop pipeline for the evolution of polymers. (b) Active-learning-guided discovery of copo-Iymer ${}^{19}$ F MRI agents. Reproduced with permission from Ref. [264]. Copyright 2021, American Chemical Society. (c) Closed-loop design-build-test-learn process for the design of polymer-protein hybrids. Reproduced with permission from Ref. [266]. Copyright 2022, John Wiley & Sons (color online).

图14(a)用于聚合物进化的机器学习辅助设计 - 构建 - 测试 - 学习闭环流程。(b)主动学习引导发现共聚物${}^{19}$ F MRI造影剂。经许可转载自参考文献[264]。版权所有2021，美国化学学会。(c)聚合物 - 蛋白质杂化物设计的闭环设计 - 构建 - 测试 - 学习过程。经许可转载自参考文献[266]。版权所有2022，约翰威立父子公司(在线彩色版)。

4.5 AI Application in biological macromolecules

4.5 人工智能在生物大分子中的应用

As mentioned previously, synthetic polymers are highly heterogeneous whose structural information is hard to encode. By contrast, biological macromolecules (such as nucleic acids and proteins) are highly programmable, which provides an exciting arena for AI application. All the details about their structures and functions are conveniently encoded in sequences and facilely manipulated by a set of biochemical tools. The structure-property relationship can thus be deduced from the mapping between sequence and function. Consequently, molecular engineering of biopoly-mers is usually accomplished with precise sequence variation. The astronomically large sequence space inevitably brings in unparalleled complexity and “the curse of dimensionality” in research and engineering. In view of the enormous biological data accumulated over the past decades (e.g., PDB, UniProt, UniClust, BFD), it is an ideal scenario for the use of AI. To date, the application of AI has already transformed many fields of bio-macromolecular research, particularly in protein sciences such as protein structure prediction, de novo protein design, and protein engineering.

如前所述，合成聚合物具有高度的异质性，其结构信息难以编码。相比之下，生物大分子(如核酸和蛋白质)具有高度的可编程性，这为人工智能的应用提供了一个令人兴奋的领域。关于它们结构和功能的所有细节都方便地编码在序列中，并可以通过一套生化工具轻松操作。因此，可以从序列与功能之间的映射推导出结构-性质关系。因此，生物聚合物的分子工程通常通过精确的序列变异来完成。巨大的序列空间不可避免地在研究和工程中带来了无与伦比的复杂性和“维度诅咒”。鉴于过去几十年积累的大量生物数据(例如，蛋白质数据银行、通用蛋白质数据库、通用聚类数据库、BFD)，这是使用人工智能的理想场景。迄今为止，人工智能的应用已经改变了生物大分子研究的许多领域，特别是在蛋白质科学领域，如蛋白质结构预测、从头蛋白质设计和蛋白质工程。

Proteins perform functions through their native structures. The Anfinsen’s dogma postulates that the native structure of one protein is determined only by the amino acid sequence as the thermodynamically most stable structure. The structure prediction thus composes the “protein folding problem”. Classically, this is accomplished by developing a reliable energy function and efficient conformational sampling protocol, as exemplified by the Rosetta software. The advent of AI-based methods pushed the structural modeling quality to approach that of experimental accuracy, resulting in a 1000- fold increase in structural data [268]. AlphaFold2 [9] and RoseTTAFold [269] learn evolutionary information from multiple sequence alignments. The use of protein language models such as ESMFold overcomes the limitation to generalize across protein families and facilitated atomic level prediction from single sequences [270]. While certain limitations remain, such as overpresentation of proteins in spite of missing features (cofactors, post-translational modifications, partners), insensitiveness to mutations, and incapability of generating dynamic ensembles, protein structure prediction seems a largely solved challenge. Predicting function from sequence using ML has also been demonstrated by assigning the enzyme commission (EC) number for a given sequence [271]. The availability of more structures further allows genome mining based on structures using tools like Foldseek [272]. These tools greatly expand our knowledge about proteins.

蛋白质通过其天然结构发挥功能。安芬森法则假设，一种蛋白质的天然结构仅由氨基酸序列决定，是热力学上最稳定的结构。因此，结构预测构成了“蛋白质折叠问题”。传统上，这是通过开发可靠的能量函数和有效的构象采样协议来实现的，如Rosetta软件所示。基于人工智能的方法的出现将结构建模质量提高到接近实验精度的水平，导致结构数据增加了1000倍[268]。AlphaFold2[9]和RoseTTAFold[269]从多序列比对中学习进化信息。使用诸如ESMFold等蛋白质语言模型克服了跨蛋白质家族进行泛化的限制，并促进了从单序列进行原子水平的预测[270]。虽然仍然存在某些限制，例如尽管存在缺失特征(辅因子、翻译后修饰、伴侣)但蛋白质的过度呈现、对突变不敏感以及无法生成动态集合，但蛋白质结构预测似乎在很大程度上是一个已解决的挑战。使用机器学习从序列预测功能也已通过为给定序列分配酶委员会(EC)编号得到证明[271]。更多结构的可用性进一步允许使用Foldseek等工具基于结构进行基因组挖掘[272]。这些工具极大地扩展了我们对蛋白质的了解。

Figure 15 (a) Closed-loop optimization of GPx activity of the heteropolypeptides via high throughput synthesis and machine learning. (b) Structure of the seven selected organohalides for heteropolypeptides library generation and aim of optimization. (c) GPx-like activity of RHPs in each iteration via random searching (blue) or Bayesian optimization (red). (d) Data validation within a plate $\left( {n = 8}\right)$ and between two different plates. RHPs with low (lanes 1-3) and high (lanes 4-7) GPx-like activities from the database were selected for validation. The dots on the right and left side in each lane represent the results from different plates. The black central lines and error bars in each lane represent the mean and s.d. The coloured line in each lane is the original activity of the RHP from the database.Reproduced with permission from Ref. [267]. Copyright 2023, Nature Publishing Group (color online).

图15(a)通过高通量合成和机器学习对杂多肽的谷胱甘肽过氧化物酶(GPx)活性进行闭环优化。(b)用于生成杂多肽文库和优化目标的七种选定有机卤化物的结构。(c)通过随机搜索(蓝色)或贝叶斯优化(红色)在每次迭代中杂合重组蛋白(RHP)的类GPx活性。(d)在一个平板$\left( {n = 8}\right)$内以及两个不同平板之间的数据验证。从数据库中选择具有低(泳道1 - 3)和高(泳道4 - 7)类GPx活性的RHP进行验证。每个泳道左右两侧的点代表来自不同平板的结果。每个泳道中的黑色中心线和误差条代表平均值和标准差。每个泳道中的彩色线是数据库中RHP的原始活性。经参考文献[267]许可转载。版权所有2023，自然出版集团(在线彩色版)。

The “inverse folding problem” of protein design aims at finding amino acid sequences that fold selectively into a desired “target” structure. More broadly, de novo protein design focuses on generating structures new to our knowledge or accomplishing functions (e.g., binding, fluorescence, catalysis) new to the scaffold. Currently, there are mainly two approaches to design, namely, data-driven and physics-inspired. The former relies on sequence features that can be extracted and leveraged by various neural network-based generative models, such as UniRep [273], ProGen [274], ESM-1b [275], ProViz [276], ProtTrans [277], and Protein-BERT [278], for structure or sequence generation. The latter combines free energy calculation, binding affinity calculation, or conformational entropy estimation with sequence variation for de novo design toward a target structure/function. It includes force-field-based methods like Rosetta [279] and FoldX [280] and ML-based tools like ABACUS [281] and ProteinMPNN [282]. To evaluate the designability of a protein fold, a backbone centered energy function of neural network, SCUBA [283], was developed. When the target structure is partially/fully absent, hallucination protocols can be employed based on trDesign [284], RFdiffusion [285], and Chroma [286]. In addition, it is also possible to perform controllable generation of proteins directly at the sequence level. The convergence of these two complementary approaches has proven synergistic and powerful in achieving hard goals such as de novo enzyme design [287,288]. The problem of these methods is the relatively low success rate, which mandates labor-intensive screening of hundreds to thousands of designs for validation. This challenge may be ameliorated by high-throughput robotic automation. While some of the obtained structures can agree precisely with the design at the atomic level, it is often difficult to gain the desired function with high activity as designed, which necessitates further rounds of directed evolution.

蛋白质设计的“反向折叠问题”旨在找到能够选择性折叠成所需“目标”结构的氨基酸序列。更广泛地说，从头蛋白质设计专注于生成我们所知的新结构或实现支架上新的功能(例如，结合、荧光、催化)。目前，主要有两种设计方法，即数据驱动和物理启发。前者依赖于各种基于神经网络的生成模型可以提取和利用的序列特征，如UniRep [273]、ProGen [274]、ESM-1b [275]、ProViz [276]、ProtTrans [277]和Protein-BERT [278]，用于结构或序列生成。后者将自由能计算、结合亲和力计算或构象熵估计与序列变异相结合，以针对目标结构/功能进行从头设计。它包括基于力场的方法，如Rosetta [279]和FoldX [280]，以及基于机器学习的工具，如ABACUS [281]和ProteinMPNN [282]。为了评估蛋白质折叠的可设计性，开发了一种以骨架为中心的神经网络能量函数SCUBA [283]。当目标结构部分/完全缺失时，可以基于trDesign [284]、RFdiffusion [285]和Chroma [286]采用幻觉协议。此外，也可以直接在序列水平上进行蛋白质的可控生成。这两种互补方法的融合已被证明在实现诸如从头酶设计[287,288]等艰巨目标方面具有协同作用且强大。这些方法的问题是成功率相对较低，这就需要对数百到数千个设计进行劳动密集型筛选以进行验证。高通量机器人自动化可能会改善这一挑战。虽然获得的一些结构在原子水平上可以与设计精确吻合，但通常很难获得设计中具有高活性的所需功能，这就需要进一步进行多轮定向进化。

Directed evolution comprises two steps: library generation and property screening. Traditionally, directed evolution takes an uphill hike on the protein fitness landscape by accumulating beneficial mutations over rounds of mutation/ screening. With high-throughput sequencing techniques and low-cost assay methods, the information about otherwise discarded suboptimal mutants can also be used to train ML models to capture the sequence-function relationship. When meaningful features are included in the representations, simple ML models such as linear regression or shallow neural network could work well, especially for those with highly correlated local mutations [289-292]. State-of-the-art protein language models can be pre-trained on sequences from all protein families and fine-tuned with multiple sequence alignments of homologues so as to be more task-specific [275]. Notably, the limited number of experimental assays (often <100) presents a considerable challenge for high-accuracy prediction using ML models. To cope with the “low-N” scenario or even enable high-accuracy zero-shot predictions, one could combine assay-labeled data and ML models trained under different contexts (e.g., probabilistic context, evolutionary context, structural-aware context) [293]. Such fitness predictors can navigate through the enormous fitness landscape by strategic virtual screening or by steered generative models [294]. To ensure broad landscape coverage with carefully chosen premium designs, one can use either a straightforward greedy algorithm or Beam search or Bayesian optimization to gain an acceptable trade-off between exploitation and exploration. It is not surprising to see that AI has already contributed considerably to directed evolution. But, protein complexity still demands heavy wet-lab experiments for directed evolution.

定向进化包括两个步骤:文库生成和性质筛选。传统上，定向进化通过在多轮突变/筛选中积累有益突变，在蛋白质适应性景观上艰难前行。借助高通量测序技术和低成本检测方法，关于原本被丢弃的次优突变体的信息也可用于训练机器学习模型，以捕捉序列-功能关系。当表示中包含有意义的特征时，简单的机器学习模型，如线性回归或浅层神经网络，可能会表现良好，特别是对于那些具有高度相关局部突变的情况[289 - 292]。最先进的蛋白质语言模型可以在来自所有蛋白质家族的序列上进行预训练，并通过同源物的多序列比对进行微调，以便更具任务特异性[275]。值得注意的是，有限数量的实验检测(通常<100)对使用机器学习模型进行高精度预测提出了相当大的挑战。为了应对“低N”情况，甚至实现高精度的零样本预测，可以将检测标记数据与在不同背景下训练的机器学习模型(例如，概率背景、进化背景、结构感知背景)相结合[293]。这种适应性预测器可以通过策略性虚拟筛选或引导生成模型在巨大的适应性景观中导航[294]。为了确保通过精心选择的优质设计实现广泛的景观覆盖，可以使用直接的贪心算法、束搜索或贝叶斯优化，以在利用和探索之间获得可接受的权衡。人工智能已经对定向进化做出了相当大的贡献，这并不奇怪。但是，蛋白质的复杂性仍然需要大量的湿实验室实验来进行定向进化。

To reduce the workload of directed evolution, continuous evolution schemes such as PACE have been developed [295], which leads to the discovery of powerful molecular biology tools like RNA polymerases [296] and base editors [297]. However, as a platform, it relies on the cell survival for selection and can hardly be adapted to non-living circumstances. To meet the enormously diverse need of protein engineering, it is increasingly recognized that an integrated biofoundry platform combining core robotic instruments (like liquid handlers, thermocyclers, fragment analyzer, and colony pickers) and AI algorithms (for data analysis and decision making) would be indispensable to enable a closed-loop in vitro continuous evolution [298] (Figure 16). The biochemical processes for making biological macromolecules are usually robust under mild conditions, which is ideal for implementing automation. For example, Plasmid-Maker has been developed as an end-to-end pipeline for automated plasmid construction [299]; BioAutomata has been developed as a closed-loop system for microbial pathway engineering [300]. Although it remains nontrivial to adapt biofoundary to diverse assay methods, this system is highly promising in terms of high-quality data generation, acquisition, and analysis. They can be used to quickly evolve the ML model to be more and more powerful over time. The interactive interface can be in the form of an AI agent specialized in protein sciences which conveniently communicates with personnel in human language. Eventually, a paradigm shift in protein engineering is envisioned. It is only with Biofoundry that the need for speed in industry could be potentially met. Ready access to diverse enzymes shall bring yet another revolution to the synthesis of small molecules with time frame and cost superior to chemical methods.

为了减少定向进化的工作量，人们已经开发了诸如PACE之类的连续进化方案[295]，这导致了强大的分子生物学工具的发现，如RNA聚合酶[296]和碱基编辑器[297]。然而，作为一个平台，它依赖于细胞存活进行选择，很难适应非生物环境。为了满足蛋白质工程极其多样化的需求，人们越来越认识到，将核心机器人仪器(如液体处理仪、热循环仪、片段分析仪和菌落挑选仪)与AI算法(用于数据分析和决策)相结合的集成生物铸造平台对于实现体外闭环连续进化是必不可少的298。制造生物大分子的生化过程在温和条件下通常很稳定，这对于实现自动化来说是理想的。例如，Plasmid-Maker已被开发为用于自动质粒构建的端到端流程[299]；BioAutomata已被开发为用于微生物途径工程的闭环系统[300]。尽管使生物铸造适应各种检测方法仍然不容易，但该系统在高质量数据生成、获取和分析方面非常有前景。它们可用于随着时间的推移快速进化ML模型，使其越来越强大。交互界面可以采用专门从事蛋白质科学的AI智能体的形式，它可以方便地用人类语言与人员进行交流。最终，人们设想蛋白质工程将发生范式转变。只有通过生物铸造，才有可能满足工业界对速度的需求。随时获取各种酶将随着时间的推移给小分子合成带来另一场革命，其时间框架和成本优于化学方法。

The above mainly focuses on protein as the model biological macromolecule. This is also where most literature works on. In principle, the work could be similarly done on RNA molecules. There are also works on using ML for structure prediction of RNA molecules [301-303]. Poly-saccharides are an exception since they are not genetically encoded and highly heterogeneous. Hence, they behave more like synthetic polymers discussed in the previous section with collective materials properties as a functional output. Nevertheless, their chemical structures may be precisely manipulated by various enzymes like glycosynthase and glycosyl transferases, and their precise structures have great implications in cell signaling and are heavily involved in diverse biological pathways. Their synthesis and structural editing may be performed using biofoundary in a way similar to proteins [304]. Overall, AI has impacted and will continue to influence the engineering of biological macromolecules for diverse purposes, especially when aided with closed-loop automation (Figure 16).

上述内容主要以蛋白质作为模型生物大分子为重点。这也是大多数文献所关注的领域。原则上，这项工作在RNA分子上也可以类似地进行。也有一些使用ML进行RNA分子结构预测的工作[301 - 303]。多糖是个例外，因为它们不是由基因编码的，而且高度异质。因此，它们的行为更类似于上一节讨论的合成聚合物，其集体材料特性作为功能输出。然而，它们的化学结构可以通过各种酶(如糖合成酶和糖基转移酶)进行精确操纵，并在细胞信号传导中具有重要意义，并且大量参与各种生物途径。它们的合成和结构编辑可以使用生物铸造以类似于蛋白质的方式进行[304]。总体而言，AI已经并将继续影响用于各种目的的生物大分子工程，特别是在闭环自动化的辅助下(图16)。

5 Automated experimentation

5 自动化实验

For AI applications in synthetic chemistry, the generation of large-scale, high-quality chemical synthesis data is not only a crucial foundation for chemical modeling but also a vital knowledge source driving the innovation of synthetic chemistry itself. However, the approach of generating synthetic chemistry data has not undergone revolutionary changes over the past century. In current chemical experimentation, manual operations still dominate, which not only makes the synthetic exploration labor-intensive but also limits the efficiency of experiments and the reproducibility of synthesis data. The advent of autonomous synthesis platforms offers a novel strategy to address these issues. These platforms, by integrating advanced control technologies and robotic systems, are capable of precision control over the chemical synthesis process, thereby enhancing the efficiency of synthetic experiments, reducing labor input, and ensuring the accuracy and reproducibility of experimental results. Recent years have witnessed significant advancements in automated synthesis, separation, and even entire intelligent synthesis systems, providing a critical hardware engine for the paradigm shift in synthetic chemistry (Figure 17).

对于人工智能在合成化学中的应用而言，生成大规模、高质量的化学合成数据不仅是化学建模的关键基础，也是推动合成化学自身创新的重要知识来源。然而，在过去的一个世纪里，生成合成化学数据的方法并没有发生革命性的变化。在当前的化学实验中，手动操作仍然占据主导地位，这不仅使合成探索劳动强度大，还限制了实验效率和合成数据的可重复性。自主合成平台的出现为解决这些问题提供了一种新策略。这些平台通过集成先进的控制技术和机器人系统，能够对化学合成过程进行精确控制，从而提高合成实验的效率，减少人力投入，并确保实验结果的准确性和可重复性。近年来，自动化合成、分离甚至整个智能合成系统都取得了显著进展，为合成化学的范式转变提供了关键的硬件引擎(图17)。

5.1 Automated synthesis

5.1 自动化合成

Automation provides an avenue to transfer organic synthesis from a labor-intensive job to a machine-driven process [305]. The first concept of automated synthesis can be traced back to the 1960s when Merrifield and Stewart reported an automated system for solid-phase peptide synthesis [306,307]. Taking advantage of a similar strategy, DNA [308,309], RNA fragments [310], as well as polysaccharides [311] could be synthesized through an automated procedure today.

自动化为将有机合成从劳动密集型工作转变为机器驱动的过程提供了一条途径[305]。自动化合成的第一个概念可以追溯到20世纪60年代，当时Merrifield和Stewart报道了一种用于固相肽合成的自动化系统[306,307]。利用类似的策略，如今DNA[308,309]、RNA片段[310]以及多糖[311]都可以通过自动化程序进行合成。

For peptides and oligonucleotides, the synthetic protocol in their automated synthesizers is fundamentally the same for every individual molecule. In contrast, this is not the same case for general molecular synthesis. The synthesis instruments need to adjust the synthetic routes to holistic, interdependent, and multistep processes, which are mostly distinctive for each synthesis of small organic molecules. Given that automation in chemical research is rare and the commercially available systems were usually designed for specific purposes and only valid for repetitive work. In 1978, Legrand and Foucard [312] developed an automation kit for synthetic chemists. Ley’s group [313] devised a convenient and efficient prototype for evaporating, concentrating, and switching solvents in continuous flow processes and batch mode. In 2013, researchers from AbbVie Inc. developed an efficient compound-synthesis system with integrated components and automated sample-handling modules [314]. Tu et al. [315] later developed a fully automated synthesis-purification station based on the SWAVE platform and in-house developed robotics. Very recently, Ahmed’s group [316] developed a robot-assisted acoustofluidic end effector (RAEE) system consisting of a robotic arm and an acous-tofluidic end effector (Figure 18a).

对于肽和寡核苷酸，其自动化合成仪中的合成方案对于每个单独的分子来说基本相同。相比之下，一般分子合成的情况并非如此。合成仪器需要将合成路线调整为整体的、相互依存的多步过程，而这些过程对于每个小分子的合成大多是独特的。鉴于化学研究中的自动化很少见，并且市售系统通常是为特定目的设计的，仅适用于重复性工作。1978年，Legrand和Foucard[312]为合成化学家开发了一种自动化套件。Ley的团队[313]设计了一种方便高效的原型，用于在连续流动过程和间歇模式下蒸发、浓缩和切换溶剂。2013年，艾伯维公司的研究人员开发了一种具有集成组件和自动化样品处理模块的高效化合物合成系统[314]。Tu等人[315]后来基于SWAVE平台和内部开发的机器人技术开发了一个全自动合成-纯化站。最近，Ahmed的团队[316]开发了一种由机器人手臂和声流体末端执行器组成的机器人辅助声流体末端执行器(RAEE)系统(图18a)。

Figure 16 Closed-loop, in vitro continuous directed evolution enabled by AI-assisted protein design and robotic automation (color online).

图16由人工智能辅助蛋白质设计和机器人自动化实现的闭环体外连续定向进化(彩色在线)。

Figure 17 Key research directions of automated experimentation (color online).

图17自动化实验的关键研究方向(彩色在线)。

Continuous flow manufacture is widely embraced for synthesizing active pharmaceutical ingredients (APIs) and fine chemicals (Figure 18b) [317,318]. Automated platforms, such as ChemKonzert (Figure 18c) [319], enable solution-phase synthesis of diverse organic compounds. Pentelute’s lab introduced an automated flow-based system for rapid polypeptide synthesis [320]. Pfizer’s platform integrates nanomole-scale screening and micromole-scale synthesis, conducting over 1,500 experiments per ${24}\mathrm{\;h}$ [321]. Li et al. [322,323] developed Tiny Tides, a fully automated fast-flow device, achieving on-demand customized antisense phosphorodiamidate morpholino oligomers (PMOs) and high-speed synthesis of PPNAs. This high-efficiency synthesizer serves as a training data source for effective ML models guiding efficient PNA sequence design [324].

连续流制造在合成活性药物成分(API)和精细化学品方面得到广泛应用(图18b)[317,318]。自动化平台，如ChemKonzert(图18c)[319]，能够实现多种有机化合物的溶液相合成。Pentelute的实验室引入了一种基于流动的自动化系统用于快速多肽合成[320]。辉瑞公司的平台集成了纳摩尔级筛选和微摩尔级合成，每${24}\mathrm{\;h}$可进行超过1500次实验[321]。Li等人[322,323]开发了Tiny Tides，一种全自动快速流动装置，可按需定制反义磷酰胺吗啉代寡聚物(PMO)并高速合成PPNA。这种高效合成器作为有效机器学习模型的训练数据源，用于指导高效的PNA序列设计[324]。

In 2019, Cronin et al. [325] developed an autonomous compiler and robotic laboratory platform, called Chemputer (Figure 18d), to synthesize organic compounds on the basis of standardized methods descriptions. Dömling’s group [326] employed I-DOT, a positive-pressure-based low-volume dispensing technology, for fully automated synthesis of over 1,000 iminopyrrolidine-2-carboxylic acid derivatives through Ugi-3-component reaction at the nanoscale. Williams, Kappe, and colleagues [327] designed an integrated multistep reaction and real-time analysis platform for controlled synthesis of mesalazine, achieving a throughput of 1.6 g per hour. In 2021, Kim’s group [328] developed a parallel flow synthesizer enabling multiplex synthesis and optimization of compound libraries, offering rapid screening and obtaining optimal conditions for various reactions in less than one hour from 96 different conditions. Gilmore and colleagues [329] introduced an automatic radial synthesizer featuring multiple continuous flow modules arranged around a central core, enabling stable and reproducible linear and convergent syntheses without manual reconfiguration. In 2022, Jensen et al. [330] developed a continuous stirred-tank reactor (CSTR) flow platform capable of handling solids and slurries during chemical transformations, enhancing the identification of optimized reaction conditions for manufacturing process development.

2019年，Cronin等人[325]开发了一个名为Chemputer的自主编译器和机器人实验室平台(图18d)，用于根据标准化方法描述合成有机化合物。Dömling的团队[326]采用I-DOT，一种基于正压的小体积分配技术，通过纳米级的Ugi-3组分反应全自动合成了超过1000种亚氨基吡咯烷-2-羧酸衍生物。Williams、Kappe及其同事[327]设计了一个集成多步反应和实时分析平台用于美沙拉嗪的可控合成，实现了每小时产率1.6克。在2021年，Kim团队[328]开发了一种平行流合成器，能够对化合物库进行多重合成和优化，可在96种不同条件下不到一小时内快速筛选并获得各种反应的最佳条件。Gilmore及其同事[329]推出了一种自动径向合成器，其具有围绕中心核心排列的多个连续流动模块，能够实现稳定且可重复的线性和收敛合成，无需手动重新配置。2月22日，Jensen等人[330]开发了一种连续搅拌釜式反应器(CSTR)流动平台，能够在化学转化过程中处理固体和浆料，有助于确定制造工艺开发的优化反应条件。

Isolation and purification in flow chemistry can follow an ideal process where reactants enter, and pure products exit continuously. George et al. [331] demonstrated continuous artemisinin synthesis in a supercritical ${\mathrm{{CO}}}_{2}$ flow system. Multi-step reactions often require interruptions for work-ups and extractions before proceeding. Inline solid-phase extraction [332], gas-liquid, and liquid-liquid separation [333] technologies can incorporate most work-ups into a continuous process. Baranczak et al. [334] developed a fully automated platform for synthesis-purification-testing of small molecule libraries. Lee and Vilela et al. [335] reported an inline chromatographic purification automated flow synthesis platform, achieving 97%-99% purity in continuously isolating products.

流动化学中的分离和纯化可以遵循一个理想的过程，即反应物进入，纯产物连续流出。George等人[331]在超临界${\mathrm{{CO}}}_{2}$流动系统中展示了青蒿素的连续合成。多步反应通常需要在进行之前中断以进行后处理和萃取。在线固相萃取[332]、气液和液液分离[333]技术可以将大多数后处理纳入连续过程。Baranczak等人[334]开发了一个用于小分子文库合成-纯化-测试的全自动平台。Lee和Vilela等人[335]报道了一个在线色谱纯化自动流动合成平台，在连续分离产物时纯度达到97%-99%。

Figure 18 Schematics of flow chemistry-based automatic synthesis platforms. (a) RAEE system. Reproduced with permission from Ref. [316]. Copyright 2022, Nature Publishing Group. (b) Flow manufacturing. Reproduced with permission from Ref. [317]. Copyright 2017, American Chemical Society. (c) ChemKonzert system. Reproduced with permission from Ref. [319]. Copyright 2010, Pharmaceutical Sociey of Japan. (d) Chemputer system. Reproduced with permission from Ref. [325]. Copyright 2019, American Association for the Advancement of Science (color online).

图18基于流动化学的自动合成平台示意图。(a) RAEE系统。经参考文献[316]许可转载。版权所有2022，自然出版集团。(b) 流动制造。经参考文献[317]许可转载。版权所有2017，美国化学学会。(c) ChemKonzert系统。经参考文献[319]许可转载。版权所有2010，日本药学会。(d) Chemputer系统。经参考文献[325]许可转载。版权所有2019，美国科学促进会(彩色在线)。

Burke’s automated Lego-like synthesis process utilized iterative peptide coupling for Suzuki-Miyaura $\mathrm{C}\left( {\mathrm{{sp}}}^{2}\right) - \mathrm{C}\left( {\mathrm{{sp}}}^{2}\right)$ bond formation [336], creating 14 diverse small molecule classes. The approach used $N$ -methyliminodiacetic acid (MIDA) as a building block, employing a “catch-and-release” purification protocol [337]. The strategy, while incompatible with stereospecific $\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right) - \mathrm{C}\left( {\mathrm{{sp}}}^{2}\right)$ or $\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right) - \mathrm{C} \; \left( {\mathrm{{sp}}}^{3}\right)$ bond-forming reactions due to MIDA sensitivity, was recently improved with stable tetramethyl- $N$ -methylimino-diacetic acid (TIDA) boronates [338]. This advancement enabled the automated synthesis of $\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right)$ boronate building blocks and facilitated stereospecific $\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right) - \mathrm{C}$ bond formation, broadening the scope of accessible molecules [339]. Jensen and collaborators [340] pioneered microfluidic automated platforms, such as droplet-based systems for efficient reaction screening and product isolation in small-scale medicinal chemistry. They optimized Pd-catalyzed C-N coupling conditions [341] and developed an automated single-droplet screening platform for electroorganic process discovery [342,343]. In 2020, Kennedy and Stephenson et al. [344] reported an automated microfluidic platform to enable picomole scale synthesis. Recently, Jensen and Pidko et al. [345] reported a catalytic asymmetric hydrogenation of a sensitive $\beta$ -amino-ketone substrate by means of an automated microfluidic platform. Debrouwer et al. [346] reported a dual catalysis cross-electrophile coupling using oscillatory plug flow photoreactors.

Burke的自动化类乐高合成过程利用迭代肽偶联进行铃木-宫浦$\mathrm{C}\left( {\mathrm{{sp}}}^{2}\right) - \mathrm{C}\left( {\mathrm{{sp}}}^{2}\right)$键形成[336]，创建了14种不同的小分子类别。该方法使用$N$-甲基亚氨基二乙酸(MIDA)作为构建单元，采用“捕获-释放”纯化方案[337]。由于MIDA的敏感性，该策略与立体特异性$\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right) - \mathrm{C}\left( {\mathrm{{sp}}}^{2}\right)$或$\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right) - \mathrm{C} \; \left( {\mathrm{{sp}}}^{3}\right)$键形成反应不兼容，但最近通过稳定的四甲基-$N$-甲基亚氨基二乙酸(TIDA)硼酸盐[338]得到了改进。这一进展实现了$\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right)$硼酸盐构建单元的自动化合成，并促进了立体特异性$\mathrm{C}\left( {\mathrm{{sp}}}^{3}\right) - \mathrm{C}$键的形成，拓宽了可获取分子的范围[339]。Jensen及其合作者[340]开创了微流控自动化平台，如基于液滴的系统，用于小规模药物化学中的高效反应筛选和产物分离。他们优化了钯催化的C-N偶联条件[341]，并开发了一个用于电有机过程发现的自动单液滴筛选平台[342,343]。2020年，Kennedy和Stephenson等人[344]报道了一个能够实现皮摩尔规模合成的自动化微流控平台。最近，Jensen和Pidko等人[345]报道了通过自动化微流控平台对敏感的$\beta$-氨基酮底物进行催化不对称氢化。Debrouwer等人[346]报道了使用振荡柱塞流光反应器的双催化交叉亲电偶联。

Adamo et al. [347] introduced a compact continuous manufacturing platform. Bode’s group [348] developed an automated capsule-based synthesis for $N$ -heterocycles. Bode et al. [349] designed an iterative console assembling molecules from vast virtual libraries. Cronin et al. [350-352] utilized 3D printing for interconnected modules and a chemical to computer-automated design (ChemCAD) approach. They later created a portable platform for universal chemical synthesis using chemical markup language (χDL) and 3D printing. These advancements signify a transformative shift towards efficient, digitized, and automated synthetic platforms [353].

Adamo等人[347]引入了一个紧凑的连续制造平台。Bode的团队[348]开发了一种用于$N$-杂环的基于胶囊的自动化合成。Bode等人[349]设计了一个从大量虚拟文库中组装分子的迭代控制台。Cronin等人[350-352]将3D打印用于相互连接的模块和化学到计算机自动化设计(ChemCAD)方法。他们后来使用化学标记语言(χDL)和3D打印创建了一个用于通用化学合成的便携式平台。这些进展标志着向高效、数字化和自动化合成平台的变革性转变[353]。

Various research groups have explored the concept of a “cloud lab” for remote operation of self-optimizing systems. In this regard, Poliakoff’s group [354] demonstrated a remote-operated system. Ley’s group [355] introduced Ley Lab in 2016, an Internet-based software allowing global monitoring and control of chemical reactions. Aspuru-Guzik and colleagues [356] developed ChemOS in 2018, a portable framework employing AI, sensors, and robotics for closed-loop systems. Cronin’s Chemputer translated reported procedures into automatable steps using NLP and $\chi$ DL [325,357]. Zhu’s materials acceleration operation system (MAOS) [358] in 2020 enabled intelligent robotics for material synthesis with AI-controlled quality assurance, accessible through VR-robot interaction. Cooper’s 2020 robo-chemist [359], driven by a Bayesian algorithm, autonomously conducted 688 reactions over eight days. Jiang’s AI-Chemist can autonomously extract literature and propose experimental plans from a cloud database [360]. In 2023, Gomes et al. [7] developed a system called Coscientist which is an artificial intelligence system driven by GPT-4 that autonomously designs, plans, and performs complex experiments.

多个研究团队探索了用于远程操作自我优化系统的“云实验室”概念。在这方面，波利亚科夫团队[354]展示了一个远程操作系统。莱伊团队[355]在2016年推出了莱伊实验室，这是一款基于互联网的软件，可对化学反应进行全球监测和控制。阿斯普鲁-古兹克及其同事[356]在2018年开发了ChemOS，这是一个用于闭环系统的便携式框架，采用了人工智能、传感器和机器人技术。克罗宁的Chemputer利用自然语言处理和$\chi$深度学习[325,357]将报告的程序转化为可自动化的步骤。朱的材料加速操作系统(MAOS)[358]在2020年实现了用于材料合成的智能机器人技术，并具备人工智能控制的质量保证功能，可通过虚拟现实与机器人的交互进行操作。库珀在2020年开发的机器人化学家[359]由贝叶斯算法驱动，在八天内自主进行了688次反应。江的人工智能化学家可以从云数据库中自动提取文献并提出实验计划[360]。2023年，戈麦斯等人[7]开发了一个名为Coscientist的系统，这是一个由GPT-4驱动的人工智能系统，能够自主设计、规划和执行复杂实验。

5.2 Automated work-up, isolation and purification

5.2 自动化后处理、分离和纯化

Automated work-up, separation, and purification platforms are integral components of laboratory automation. In this regard, Ley et al. [361] has done significant contributions, whose works have been comprehensively reviewed by their own review. For instance, they have implemented machine vision automation for extraction operations [362], online solvent flash evaporation devices [313], optimized chromatographic separations [363], and automated filtration [364]. In terms of automating chemical laboratories and integrating ML and deep learning, the Cronin research group has achieved remarkable progress for automated automated synthesis machine [325,365]. Their system’s implementation relies on computer-controlled pumps. These pumps inject reactants into reaction flasks. Reaction work-up, including extraction, column chromatography, and rotary evaporation, is also integral to the system. These operations are achieved by transferring liquid reactants through a complex pipeline system using pumps. Spectroscopic detection methods, such as infrared spectroscopy and nuclear magnetic resonance spectroscopy, are also integrated. ML algorithms are employed to interpret these spectra, obtaining reaction information, which is then fed back into the system to achieve a closed-loop optimization.

自动化后处理、分离和纯化平台是实验室自动化的重要组成部分。在这方面，Ley等人[361]做出了重大贡献，他们的工作已在其自身的综述中得到全面回顾。例如，他们实现了用于萃取操作的机器视觉自动化[362]、在线溶剂闪蒸装置[313]、优化的色谱分离[363]以及自动化过滤[364]。在化学实验室自动化以及机器学习和深度学习的整合方面，Cronin研究小组在自动化合成机器方面取得了显著进展[325,365]。他们系统的实现依赖于计算机控制的泵。这些泵将反应物注入反应烧瓶。包括萃取、柱色谱和旋转蒸发在内的反应后处理也是该系统的一部分。这些操作通过使用泵通过复杂的管道系统输送液体反应物来实现。还集成了光谱检测方法，如红外光谱和核磁共振光谱。采用机器学习算法来解释这些光谱，获取反应信息，然后将其反馈到系统中以实现闭环优化。

In chromatographic analysis and preparation, Kassel et al.’s PrepLCMS [366], a pioneering mass spectrometry-based system, automates the purification of substantial compound quantities. Koppitz et al.’s LC/MS-based system efficiently processes 100-200 compounds daily [367], ensuring high purity and yield. Ilg et al. [368] introduced a high-throughput high performance liquid chromatography/ mass spectrometry (HPLC/MS) platform, incorporating Covaris technology for sample preparation, automated ali-quotation in fractionation, and a novel evaporation technique combining freeze-drying, enhancing purification efficiency. Recently, Mo et al. [369] developed an automated thin layer chromatography (TLC) platform for high-throughput data collection, subsequently using ML methods to predict the retardation factor (Rf) of compounds. The trained ML model can accurately predict the Rf value curves of organic compounds under different solvent combinations, providing general guidance for purification condition selection. Additionally, they have also developed a QGeoGNN-based model for predicting optimal HPLC separation conditions for chiral enantiomers, significantly reducing trial-and-error costs [370].

在色谱分析和制备中，Kassel等人的PrepLCMS[366]是一个开创性的基于质谱的系统，可实现大量化合物的自动化纯化。Koppitz等人基于液相色谱/质谱的系统每天可高效处理100 - 200种化合物[367]，确保高纯度和高收率。Ilg等人[368]引入了一个高通量高效液相色谱/质谱(HPLC/MS)平台，采用Covaris技术进行样品制备、分馏中的自动等分以及一种结合冷冻干燥的新型蒸发技术，提高了纯化效率。最近，Mo等人[369]开发了一个用于高通量数据收集的自动化薄层色谱(TLC)平台，随后使用机器学习方法预测化合物的比移值(Rf)。经过训练的机器学习模型可以准确预测不同溶剂组合下有机化合物的Rf值曲线，为纯化条件选择提供通用指导。此外，他们还开发了一种基于QGeoGNN的模型，用于预测手性对映体的最佳HPLC分离条件，显著降低了试错成本[370]。

5.3 Integration of AI with robotic systems

5.3人工智能与机器人系统的集成

Discovering new reactions is unpredictable and laborious. Suboptimal initial conditions, especially in micro/nanoscale, may lead to overlooked trace products. In this regard, the integration of AI with robotic system can provide an effective strategy. However, it should be noted that applying ML to navigate new chemical space is underexplored due to the challenge of assessing reactivity in unknown reactions with unpredictable products compared with optimizing conditions for known target compounds [371,372]. Deconvolution algorithms, for instance, can help identify novel products [373,374]. Cronin’s group [365] demonstrated that a synthetic robot controlled by SVM algorithm significantly accelerated organic reaction discovery. The liquid-handling robot selected reactants from a pool, with real-time analytics monitoring reactions. ML built a chemical space model, recommending experiments and controlling the robot. The system outperformed manual processes, predicting the reactivity of 1,000 combinations with over 80% accuracy. Zahrt and colleagues [375] applied ML to guide electrochemical reaction discovery, developing a molecular representation for general models and successfully predicting new reactions’ competency. These studies showcase AI-driven chemical robots advancing reaction space exploration. Recently, research groups have also applied reinforcement learning for automated mechanism discovery, bypassing exhaustive screening [376,377]. An agent constructs efficient reaction pathways by selecting actions (elementary steps) with varying rewards. This approach holds promise for efficient reaction network exploration, requiring first-principles or semi-empirical evaluations.

发现新反应是不可预测且费力的。次优的初始条件，尤其是在微观/纳米尺度下，可能会导致痕量产物被忽视。在这方面，人工智能与机器人系统的集成可以提供一种有效的策略。然而，应该注意的是，由于与优化已知目标化合物的条件相比，在具有不可预测产物的未知反应中评估反应性具有挑战性，因此将机器学习应用于探索新的化学空间尚未得到充分探索[371,372]。例如，反卷积算法可以帮助识别新产物[373,374]。克罗宁的团队[365]证明，由支持向量机算法控制的合成机器人显著加速了有机反应的发现。液体处理机器人从一组反应物中选择反应物，并通过实时分析监测反应。机器学习建立了一个化学空间模型，推荐实验并控制机器人。该系统优于手动操作流程，以超过80%的准确率预测了1000种组合的反应性。扎尔特及其同事[375]应用机器学习来指导电化学反应的发现，为通用模型开发了一种分子表示法，并成功预测了新反应的能力。这些研究展示了人工智能驱动的化学机器人在推进反应空间探索方面的作用。最近，研究团队也将强化学习应用于自动机理发现，绕过了详尽的筛选[376,377]。一个智能体通过选择具有不同奖励的行动(基本步骤)来构建高效的反应路径。这种方法有望实现高效的反应网络探索，这需要第一性原理或半经验评估。

6 Challenges and perspective

6挑战与展望

The burgeoning field of AI in organic and polymer synthesis presents a transformative potential for scientific discovery. However, this promise is contingent on overcoming a series of challenges that currently impede its full realization. This section delves into these critical issues, offering a succinct yet comprehensive overview of the challenges faced by AI applications in synthetic chemistry as well as potential solutions.

人工智能在有机和聚合物合成领域的蓬勃发展为科学发现带来了变革潜力。然而，这一前景取决于克服一系列目前阻碍其全面实现的挑战。本节深入探讨这些关键问题，简要而全面地概述了人工智能在合成化学应用中面临的挑战以及潜在解决方案。

6.1 Data

6.1 数据

In synthetic chemistry, particularly in organic and polymer synthesis, the role of data is foundational for the successful AI application [378]. The main challenges associated with data in this field include issues of quantity, quality, standardization, and accessibility [229,379,380]. Generating sufficient, high-quality data is a complex endeavor, limited by the intricate and time-consuming nature of chemical experiments. The quality of data, essential for the training and performance of AI models, is frequently compromised by variations in experimental conditions, disparate practices among researchers, and inherent biases, leading to significant inconsistencies [381-383]. These variations and the lack of detailed reaction conditions in public databases undermine the reliability of data for AI applications. Moreover, the absence of standardized data formats complicates the compatibility and comparability across different systems, hindering the efficient training of AI models and their application to varied tasks. Data accessibility is further challenged by legal, technical, and proprietary barriers that restrict the use of data, making it difficult for researchers to obtain and utilize the information needed for their work.

在合成化学中，尤其是在有机和聚合物合成领域，数据的作用是人工智能成功应用的基础[378]。该领域与数据相关的主要挑战包括数量、质量、标准化和可获取性问题[229,379,380]。生成足够的高质量数据是一项复杂的工作，受到化学实验复杂且耗时特性的限制。对于人工智能模型的训练和性能至关重要的数据质量，经常因实验条件的变化、研究人员之间不同的做法以及固有偏差而受到影响，导致显著的不一致性[381 - 383]。这些变化以及公共数据库中缺乏详细的反应条件，破坏了人工智能应用数据的可靠性。此外，缺乏标准化的数据格式使不同系统之间的兼容性和可比性变得复杂，阻碍了人工智能模型的高效训练及其在各种任务中的应用。数据的可获取性还受到法律、技术和专有障碍的进一步挑战，这些障碍限制了数据的使用，使得研究人员难以获取和利用其工作所需的信息。

Addressing the limitations around data in synthetic chemistry necessitates a multifaceted approach that integrates the establishment of open data principles with advanced AI-assisted data management techniques. The adoption of FAIR principles—ensuring data is Findable, Accessible, Interoperable, and Reusable-is critical for improving data quality, standardization, and accessibility [384]. These principles support the creation of a standardized data management framework that facilitates the sharing and reuse of data across the scientific community. Additionally, leveraging AI for automated data extraction and processing offers a powerful solution to enhance the efficiency and accuracy of data collection [385,386]. This involves the use of advanced natural language processing and large language models for scraping, mining, and extracting valuable information from a plethora of sources including chemical literature, reaction databases, and experimental records. The key to harnessing these technologies lies in their ability to process and analyze vast amounts of data, translating them into actionable insights that can drive research forward. However, ensuring the reliability of the extracted data is crucial, necessitating careful validation and verification processes. Moreover, fostering an open data community, grounded in the principles of collaboration and shared resources, is essential for overcoming the barriers of data accessibility and standardization. Such a community would serve as a hub for aggregating, refining, and sharing data, thereby facilitating a more collaborative, efficient, and innovative research environment [35]. Together, these strategies offer a comprehensive blueprint for addressing the challenges posed by data limitations in synthetic chemistry, paving the way for enhanced AI applications and scientific discovery.

解决合成化学中数据方面的局限性需要采取多方面的方法，将开放数据原则的建立与先进的人工智能辅助数据管理技术相结合。采用FAIR原则(确保数据是可查找、可获取、可互操作和可重用的)对于提高数据质量、标准化和可获取性至关重要[384]。这些原则支持创建一个标准化的数据管理框架，促进整个科学界的数据共享和重用化。此外，利用人工智能进行自动数据提取和处理提供了一个强大的解决方案，以提高数据收集的效率和准确性[385,386]。这涉及使用先进的自然语言处理和大语言模型，从包括化学文献、反应数据库和实验记录在内的大量来源中抓取、挖掘和提取有价值的信息。利用这些技术的关键在于它们处理和分析大量数据的能力，将其转化为可推动研究前进的可操作见解。然而，确保提取数据的可靠性至关重要，需要仔细的验证和核实过程。此外，建立一个基于合作和共享资源原则的开放数据社区，对于克服数据可获取性和标准化的障碍至关重要。这样的社区将成为聚合、完善和共享数据的中心，从而促进更具协作性、高效性和创新性的研究环境[35]。总之，这些策略为应对合成化学中数据局限性带来的挑战提供了一个全面的蓝图，为增强人工智能应用和科学发现铺平了道路。

6.2 Encoding

6.2 编码

For the digital representation of synthetic chemistry, three core challenges are prominently identified: universality, interpretability, and the representation of the stochastic nature of polymer structures. The complexity and diversity of chemical data, spanning a wide spectrum from molecular structures to reaction conditions, necessitate distinct representation approaches for each type, complicating the quest for universality. The issue of diversity, encompassing various modalities such as texts, images, and tables, adds another layer of complexity in standardizing chemical information. Further complicating this pursuit is the fact that various laboratories and researchers often employ their customized methods for data recording and representation. These personalized approaches create significant hurdles in achieving a universal standard for chemical data across different formats and sources. The challenge of interpretability arises from the need to encode chemical insights in a way that is comprehensive to computational models and intelligible to human researchers. This includes difficulties in conveying complex chemical phenomena and the inherent tension in designing models that combine high accuracy with ease of understanding, emphasizing the trade-off where increased predictive performance often diminishes transparency. Additionally, accurately capturing the stochastic nature of polymers, characterized by their varied molecular weights and structural configurations, presents a unique challenge. The properties of polymers are heavily influenced by their molecular diversity, requiring nuanced and precise encoding strategies to capture the essential characteristics that dictate their behavior and functionality. These challenges collectively underscore the complexities of developing effective encoding systems in synthetic chemistry, aimed at bridging the gap between the intricate chemical phenomena and their computational representations.

对于合成化学的数字表示，突出识别出三个核心挑战:通用性、可解释性以及聚合物结构随机性质的表示。化学数据的复杂性和多样性，涵盖从分子结构到反应条件的广泛范围，使得每种类型都需要不同的表示方法，这使得追求通用性变得复杂。多样性问题，包括文本、图像和表格等各种模态，在标准化化学信息方面又增加了一层复杂性。进一步使这一追求复杂化的是，各个实验室和研究人员经常采用他们定制的数据记录和表示方法。这些个性化方法在实现跨不同格式和来源的化学数据通用标准方面造成了重大障碍。可解释性挑战源于需要以一种对计算模型全面且对人类研究人员易懂的方式对化学见解进行编码。这包括在传达复杂化学现象方面的困难，以及在设计结合高精度与易理解性的模型时存在的内在矛盾，强调了在提高预测性能时往往会降低透明度的权衡。此外，准确捕捉聚合物的随机性质，其特征在于不同的分子量和结构构型，提出了一个独特的挑战。聚合物的性质受到其分子多样性的严重影响，需要细致而精确的编码策略来捕捉决定其行为和功能的基本特征。这些挑战共同凸显了在合成化学中开发有效编码系统的复杂性，旨在弥合复杂化学现象与其计算表示之间的差距。

Advancing encoding techniques in synthetic chemistry can be approached from the following angles. Leveraging multimodal learning methods and large Transformer-based models such as ChemBERTa [387], MoLFormer [388], and ChemGPT [389] to integrate chemical data from diverse modalities including texts, images, and tables, could pave the way towards a unified representation system. This effort may also involve standardizing chemical information through the creation of universal datasets and the application of intelligent algorithms to address the challenges of non-standardization. On the interpretability front, enhancing models to combine high accuracy with ease of understanding is crucial. A representative example is the ASO descriptor designed by Denmark et al. [187], which finely depicts the three-dimensional structure of chiral molecules from the perspective of space filling. Additionally, symbolic regression represents a valuable method that could uncover relationships with clear analytical expressions, offering new ways to interpret complex chemical data [390]. For the representation of the stochastic nature of polymers, it requires models that can encapsulate the diversity in molecular weights and structural configurations. Specialized polymer representation models that consider dispersity and monomer sequence arrangements could more precisely predict polymer properties [214]. Exploring computational models for topological structures, such as branched polymers, might also improve the accuracy of property predictions and expand the models’ applicability. Through these strategies, the goal is to effectively bridge the gap between the complexity of chemical phenomena and their digital representation, facilitating AI applications for chemical understanding and innovation.

可以从以下角度推进合成化学中的编码技术。利用多模态学习方法和基于Transformer的大型模型，如ChemBERTa [387]、MoLFormer [388] 和ChemGPT [389]，来整合来自文本、图像和表格等不同模态的化学数据，可能为统一表示系统铺平道路。这项工作还可能涉及通过创建通用数据集和应用智能算法来标准化化学信息，以应对非标准化挑战。在可解释性方面，增强模型以结合高精度与易理解性至关重要。一个有代表性的例子是丹麦等人 [187] 设计的ASO描述符，它从空间填充的角度精细地描绘了手性分子的三维结构。此外，符号回归是一种有价值的方法，可以揭示具有清晰解析表达式的关系，为解释复杂化学数据提供新方法 [390]。对于聚合物随机性质的表示，需要能够封装分子量和结构构型多样性的模型。考虑分散性和单体序列排列的专门聚合物表示模型可以更精确地预测聚合物性质 [214]。探索用于拓扑结构(如支化聚合物)的计算模型，也可能提高性质预测的准确性并扩大模型的适用性。通过这些策略，目标是有效弥合化学现象的复杂性与其数字表示之间的差距，促进人工智能在化学理解和创新方面的应用。

6.3 Model availability

6.3 模型可用性

In synthetic chemistry, the field faces the challenge of model availability due to its diverse chemical dimensions and highly individualized application scenarios. Related AI researches often narrow the focus to specific synthetic targets, utilizing customized datasets of limited size. This specialized approach to developing and implementing AI models in synthetic chemistry is not yet fully mature. Although certain AI models demonstrate significant potential, the majority presented in research papers typically provide only a GitHub link with minimal annotations. This mode of sharing, while enabling the replication of research, lacks in offering user-friendly software, platforms, or sufficient documentation and user guides, rendering it difficult for chemists without computer science expertise to effectively utilize these models. Additionally, most model developments prioritize the verification of scientific hypotheses over the consideration of the models’ applicability from the users’ perspective. Even successful model implementations may not meet the specific needs of synthetic chemists for particular molecules or reactions. The efficiency and accessibility of the encoding process also pose notable challenges. Many of the current models require the use of specialized quantum chemistry software and significant computational resources, further complicating matters for experimental synthetic chemists. Thus, making AI technology conveniently and efficiently usable for experimental chemists is essential for the progress of AI in synthetic chemistry.

在合成化学领域，由于其化学维度多样且应用场景高度个性化，面临着模型可用性的挑战。相关的人工智能研究往往将重点局限于特定的合成目标，使用规模有限的定制数据集。这种在合成化学中开发和实施人工智能模型的专门方法尚未完全成熟。尽管某些人工智能模型显示出巨大潜力，但研究论文中呈现的大多数模型通常只提供一个带有最少注释的GitHub链接。这种共享模式虽然能实现研究的复制，但缺乏用户友好的软件、平台或足够的文档及用户指南，使得没有计算机科学专业知识的化学家难以有效利用这些模型。此外，大多数模型开发优先考虑科学假设的验证，而不是从用户角度考虑模型的适用性。即使成功实施的模型也可能无法满足合成化学家对特定分子或反应的具体需求。编码过程的效率和可及性也带来了显著挑战。当前许多模型需要使用专门的量子化学软件和大量计算资源，这让实验合成化学家的情况更加复杂。因此，使人工智能技术方便且高效地供实验化学家使用，对于人工智能在合成化学中的发展至关重要。

To tackle the issue of availability, democratizing AI becomes a critical step towards technological advancement, essential for fostering scientific innovation of AI-assisted synthetic advancement. This democratization process aims to make AI tools, algorithms, and software more accessible and user-friendly for synthetic chemists, particularly those without a background in computer science. Developing AI software that aligns with the needs and experiences of chemists, moving away from complex code repositories to tools characterized by intuitive data input, clear result displays, and simple operation procedures, can significantly lower the barriers to AI application, thereby improving its impact in synthetic chemistry. Moreover, the transparency and chemical interpretability of AI tools are crucial; they should not only provide accurate predictions but also clearly explain their decision-making processes to users, building trust and promoting positive interactions between chemists and AI. In addition, allowing chemists to contribute to the AI modeling processes can lead to more meaningful predictions and ensure that AI-assisted experimental designs are closely aligned with real-world scenarios. The democratization of AI in synthetic chemistry is not just about making AI more accessible; it is about creating a more collaborative and innovative environment where AI and synthetic chemistry complement each other.

为了解决可用性问题，人工智能的普及成为技术进步的关键一步，这对于促进人工智能辅助合成进步的科学创新至关重要。这一普及过程旨在使人工智能工具、算法和软件对合成化学家，特别是那些没有计算机科学背景的化学家来说更容易获取且更用户友好。开发符合化学家需求和经验的人工智能软件，从复杂的代码库转向具有直观数据输入、清晰结果显示和简单操作程序特点的工具，可以显著降低人工智能应用的障碍，从而提高其在合成化学中的影响力。此外，人工智能工具的透明度和化学可解释性至关重要；它们不仅应提供准确的预测，还应向用户清楚解释其决策过程，建立信任并促进化学家和人工智能之间的积极互动。此外，允许化学家参与人工智能建模过程可以带来更有意义的预测，并确保人工智能辅助的实验设计与实际情况紧密结合。合成化学中人工智能的普及不仅是让人工智能更容易获取；而是要创造一个更具协作性和创新性的环境，使人工智能和合成化学相互补充。

6.4 Automated experimentation

6.4 自动化实验

Automated or semi-automated platforms, utilizing robotics and data-driven algorithms, present a solution to the bottleneck in chemical synthesis. While automation is well-established in routine tasks for pharmaceuticals, it often focuses on narrow, well-defined processes. Methodologies for chemical synthesis automation, optimization, and discovery, particularly in laboratory-based research and bench-scale synthesis, face challenges from both hardware and software, as well as the high cost. For the hardware foundation, a critical issue is the integration of automated synthetic platforms into existing laboratory setups. This requires not only consideration of the physical space within synthesis labs but also the adaptability of the platform to seamlessly fit these environments. Moreover, the importance of user-friendly software interfaces and application programming interfaces (APIs) cannot be overstated. Chemists are in search of comprehensive solutions that encompass reaction monitoring, machine self-optimization, and AI/ML algorithms, all compatible with remote control capabilities. Currently, the software and APIs available fall short of supporting a fully self-driving laboratory, indicating a significant gap that needs to be bridged.

利用机器人技术和数据驱动算法的自动化或半自动化平台，为化学合成中的瓶颈问题提供了解决方案。虽然自动化在制药的常规任务中已很成熟，但它通常侧重于狭窄、定义明确的流程。化学合成自动化、优化和发现的方法，特别是在基于实验室的研究和实验室规模合成中，面临着来自硬件和软件以及高成本的挑战。对于硬件基础而言，一个关键问题是将自动化合成平台集成到现有的实验室设置中。这不仅需要考虑合成实验室内的物理空间，还需要考虑平台无缝适应这些环境的能力。此外，用户友好的软件界面和应用程序编程接口(API)的重要性再怎么强调也不为过。化学家们正在寻找涵盖反应监测、机器自我优化和人工智能/机器学习算法的综合解决方案，所有这些都要与远程控制功能兼容。目前，现有的软件和API不足以支持一个完全自动驾驶的实验室，这表明存在一个需要弥合的重大差距。

In tackling the challenges faced by automated chemical synthesis, a focused approach on both hardware and software innovations is pivotal. For hardware, the introduction of customizable, modular systems like Opentrons’ laboratory robots for liquid handling showcases a significant step towards affordability and adaptability in automation. Their open-source robots (OT-1 and OT-2), priced as low as $10,000, exemplify the move towards making sophisticated automated platforms more accessible to a broader audience, ensuring easy integration into existing lab setups without extensive modifications. On the software side, beyond democratizing AI, the incorporation of integrated AI management software holds the key to bridging the gap in automated chemical synthesis. Systems like those developed by Jensen and Jamison, utilizing MATLAB and LabVIEW, offer examples of how control systems can provide real-time monitoring and automated feedback optimization [391]. Such platforms demonstrate the potential for AI-based synthesis planning and ML algorithms to revolutionize synthesis routes, from hypothesis generation to molecule structure prediction. Together, these hardware and software advancements present a coherent strategy to overcome existing obstacles in automated chemical synthesis. By aligning the cost-effective, customizable hardware solutions with cutting-edge, integrated software platforms, the field is set to undergo a transformative shift towards more accessible, efficient, and innovative research methodologies, marking a significant leap in the application of automation technology within the chemical sciences.

在应对自动化化学合成所面临的挑战时，专注于硬件和软件创新的方法至关重要。对于硬件而言，引入可定制的模块化系统，如用于液体处理的Opentrons实验室机器人，是朝着实现自动化的可承受性和适应性迈出的重要一步。其开源机器人(OT - 1和OT - 2)，价格低至10,000美元，体现了使复杂的自动化平台更易于广大用户使用的趋势，确保能够轻松集成到现有的实验室设置中，而无需进行大量修改。在软件方面，除了让人工智能民主化之外，集成人工智能管理软件的纳入是弥合自动化化学合成差距的关键。像Jensen和Jamison开发的系统，利用MATLAB和LabVIEW，展示了控制系统如何提供实时监测和自动反馈优化[391]。这样的平台展示了基于人工智能的合成规划和机器学习算法在革新合成路线方面的潜力，从假设生成到分子结构预测。总之，这些硬件和软件的进步提出了一个连贯的策略来克服自动化化学合成中的现有障碍。通过将具有成本效益的可定制硬件解决方案与前沿的集成软件平台相结合，该领域将朝着更易于使用、高效和创新的研究方法发生变革性转变，标志着自动化技术在化学科学中的应用取得了重大飞跃。

7 Conclusions and outlook

7 结论与展望

In summary, this review discusses the applications of AI in organic and polymer synthesis in recent years, investigating the benefits and potential of the data-driven research paradigm in addressing challenges of synthetic chemistry. In organic synthesis, AI applications have made significant breakthroughs at various levels ranging from molecules to reactions: (1) Predictions of molecular thermodynamic and kinetic properties have seen a quantum leap in efficiency without sacrificing accuracy. Chemists, empowered by ML models, can now swiftly and precisely assess crucial physicochemical parameters like $\mathrm{p}{K}_{\mathrm{a}}$ , BDE, and rate constants, offering valuable insights for molecular design in synthetic chemistry. (2) The capabilities of computer-assisted synthetic planning have undergone tremendous improvement, particularly for complex molecules. Emerging AI software can more rationally, diversely, and efficiently plan multi-step synthetic routes, even rivaling human chemist designs. (3) Data-driven prediction for yield and selectivity can help chemists identify superior catalysts or reagents, providing essential AI support for rational reaction design. In polymer synthesis, AI application has also shown remarkable outcomes: (1) ML methods can establish quantitative relationship between polymer structures and properties, achieving accurate predictions and even target-oriented polymer design; (2) AI can aid and guide the design and optimization of polymerization processes, achieving end-to-end control and linking polymerization conditions directly to the products’ functionalities; (3) AI application in biological macromolecules is equally thriving, predicting structures, designing functional sequences, and even autonomously performing closed-loop continuous directed evolution for proteins, RNA, and other macromolecules. Additionally, the advancement of automated experimentation paves the way for liberating synthetic chemists, significantly improving precision and efficiency in synthesis, work-up, isolation, and purification. This, coupled with AI’s brainpower, heralds the advent of intelligent synthesis laboratories. These exciting developments demonstrate AI’s substantial contribution to synthetic chemistry, signaling the dawn of an era of intelligent synthesis.

总之，本综述讨论了近年来人工智能在有机和聚合物合成中的应用，研究了数据驱动的研究范式在应对合成化学挑战方面的优势和潜力。在有机合成中，人工智能的应用在从分子到反应的各个层面都取得了重大突破:(1)分子热力学和动力学性质的预测在不牺牲准确性的情况下效率得到了飞跃。借助机器学习模型，化学家现在可以快速准确地评估诸如$\mathrm{p}{K}_{\mathrm{a}}$、键解离能(BDE)和速率常数等关键物理化学参数，为合成化学中的分子设计提供有价值的见解。(2)计算机辅助合成规划的能力有了巨大提升，特别是对于复杂分子。新兴的人工智能软件可以更合理、多样和高效地规划多步合成路线——甚至可以与人类化学家的设计相媲美。(3)基于数据的产率和选择性预测可以帮助化学家识别优质的催化剂或试剂，为合理的反应设计提供重要的人工智能支持。在聚合物合成中，人工智能的应用也取得了显著成果:(1)机器学习方法可以建立聚合物结构与性能之间的定量关系，实现准确预测甚至是目标导向的聚合物设计；(2)人工智能可以辅助和指导聚合过程的设计和优化，实现端到端控制，并将聚合条件直接与产品功能联系起来；(3)人工智能在生物大分子中的应用同样蓬勃发展，可预测结构、设计功能序列，甚至能对蛋白质、RNA和其他大分子自主进行闭环连续定向进化。此外，自动化实验的进步为解放合成化学家铺平了道路，显著提高了合成、后处理、分离和纯化的精度和效率。这与人工智能的智慧相结合，预示着智能合成实验室的到来。这些令人兴奋的发展证明了人工智能对合成化学的重大贡献，标志着智能合成时代的曙光。

However, it is crucial to acknowledge that AI application in synthetic chemistry is still nascent, with challenges and limitations that cannot be overlooked. The quantity and quality of available open data in synthetic chemistry are far from satisfactory, lacking unbiased, large-scale datasets like ImageNet to support AI development. The digital representation of synthetic systems requires the improvements in standardization, interpretability, and applicability. Current molecular and reaction encodings lack standardized methods and deep chemical understanding, also posing challenges in encoding like the stochastic nature of polymer structures. The “black box” nature of existing AI models makes it difficult for chemists to comprehend the decision-making process, limiting the models’ capacity to provide chemical insights. Issues with model availability also hinder the model application in new synthetic systems. To overcome these challenges and truly promote the healthy, sustainable development of AI synthetic chemistry, the following actions are recommended. First, enhancing data sharing and model openness. Following the FAIR (findable, accessible, inter-operable, and reusable) principles, chemists should reshape the open data community of synthetic chemistry with advanced large models, so as to foster broader collaboration and innovation. Second, democratize AI, making the cutting-edge achievements of AI chemistry accessible to experimental chemists for frontline synthetic design. Third, focusing on software and hardware upgrades and optimization to better integrate AI technology and experimental synthetic processes. This will enable automated synthesis platforms to enter everyday laboratories and significantly enhance the efficiency and accuracy of synthetic experiments. It requires the joint efforts of chemists, computer scientists, and engineers to accelerate the process of intelligentization in synthetic chemistry.

然而，必须认识到人工智能在合成化学中的应用仍处于起步阶段，存在一些不可忽视的挑战和局限性。合成化学中可用的开放数据的数量和质量远不能令人满意，缺乏像ImageNet那样无偏差的大规模数据集来支持人工智能的发展。合成系统的数字表示需要在标准化、可解释性和适用性方面加以改进。当前的分子和反应编码缺乏标准化方法以及对化学的深入理解，在编码方面也带来了挑战，比如聚合物结构的随机性。现有人工智能模型的“黑箱”性质使得化学家难以理解其决策过程，限制了模型提供化学见解的能力。模型可用性问题也阻碍了模型在新的合成系统中的应用。为了克服这些挑战并真正推动人工智能合成化学的健康、可持续发展，建议采取以下行动。首先，加强数据共享和模型开放性。遵循FAIR(可查找、可访问、可互操作和可重用)原则——化学家应利用先进的大型模型重塑合成化学的开放数据社区，以促进更广泛的合作与创新。其次，使人工智能民主化，让实验化学家能够在前沿合成设计中使用人工智能化学的前沿成果。第三，专注于软件和硬件的升级与优化，以更好地将人工智能技术与实验合成过程相结合。这将使自动化合成平台进入日常实验室，并显著提高合成实验的效率和准确性。这需要化学家、计算机科学家和工程师共同努力，以加速合成化学的智能化进程。

With the continuous breakthroughs and development of ML and automation technologies, synthetic chemistry is undergoing a transformation from a traditional “manual” era to an “intelligent” era. In the near future, AI will play a vital role in every aspect of synthetic chemistry: (1) Molecular design. AI will use chemical databases and ML algorithms to design molecules with specific functions according to chemists’ needs. (2) Synthetic pathway planning. For a given target molecule, AI models can efficiently plan synthetic pathways and provide detailed experimental schemes. (3) Experimental execution: Integrated AI with automated synthesis platforms/robots will create intelligent synthesis laboratories capable of conducting experiments, providing real-time feedback, and automatically adjusting experimental plans for condition optimization until the target product is synthesized with high selectivity and yield. (4) Remote interaction: AI systems deployed in the cloud enable chemists to interact remotely at any time and place via mobile phones/computers. (5) The commercialization of general chemical databases, ML algorithms, and machine chemists will likely make AI and related automation technologies standard equipment in ordinary synthetic laboratories, greatly promoting the development of synthetic chemistry.

随着机器学习和自动化技术的不断突破与发展，合成化学正从传统的“手工”时代向“智能”时代转变。在不久的将来，人工智能将在合成化学的各个方面发挥至关重要的作用:(1)分子设计。人工智能将利用化学数据库和机器学习算法，根据化学家的需求设计具有特定功能的分子。(2)合成路线规划。对于给定的目标分子，人工智能模型可以高效地规划合成路线，并提供详细的实验方案。(3)实验执行:将人工智能与自动化合成平台/机器人集成，将创建能够进行实验、提供实时反馈并自动调整实验方案以优化条件的智能合成实验室，直到以高选择性和产率合成目标产物。(4)远程交互:部署在云端的人工智能系统使化学家能够通过手机/计算机在任何时间和地点进行远程交互；(5)通用化学数据库、机器学习算法和机器化学家的商业化可能会使人工智能和相关自动化技术成为普通合成实验室的标准设备，极大地推动合成化学的发展。

Acknowledgements This work was supported by the National Natural Science Foundation of China (22393890, You SL; 22393891 and 22031006, Luo S; 2203300, Pei J; 22371052, Chen M; 21991132, 21925102, 92056118, and 22331003, Zhang WB; 22331002 and 22125101, Lu H; 22071004, Mo F; 22393892 and 22071249, Liao K; 22122109 and 22271253, Hong X), the National Key R&D Program of China (2023YFF1205103, Pei J; 2020YFA0908100 and 2023YFF1204401, Zhang WB; 2022YFA1504301, Hong X), Zhejiang Provincial Natural Science Foundation of China (LDQ23B020002, Hong X), the Starry Night Science Fund of Zhejiang University Shanghai Institute for Advanced Study (SN-ZJU-SIAS-006, Hong X), the CAS Youth Interdisciplinary Team (JCTD- 2021-11, Hong X), Shenzhen Medical Research Fund (B2302037, Zhang WB), Beijing National Laboratory for Molecular Sciences (BNLMS-CXXM-202006, Zhang WB), the State Key Laboratory of Molecular Engineering of Polymers (Chen M), Haihe Laboratory of Sustainable Chemical Transformations and National Science & Technology Fundamental Resource Investigation Program of China (2023YFA1500008, Luo S).

致谢 本研究得到了中国国家自然科学基金(22393890，游书力；22393891和22031006，罗三中；2203300，裴剑锋；22371052，陈敏；21991132、21925102、92056118和22331003，张万斌；22331002和22125101，卢航；22071004，莫凡；22393892和22071249，廖矿标；22122109和22271253，洪鑫)、中国国家重点研发计划(2023YFF1205103，裴剑锋；2020YFA0908100和2023YFF1204401，张万斌；2022YFA1504301，洪鑫)、浙江省自然科学基金(LDQ23B020002，洪鑫)、浙江大学上海高等研究院星夜科学基金(SN-ZJU-SIAS-006，洪鑫)、中国科学院青年交叉团队(JCTD-2021-11，洪鑫)、深圳市医学科研基金(B2302037，张万斌)、北京分子科学国家实验室(BNLMS-CXXM-202006，张万斌)、聚合物分子工程国家重点实验室(陈敏)、海河可持续化学转化实验室以及中国国家科技基础资源调查专项(2023YFA1500008，罗三中)的资助。

Conflict of interest The authors declare no conflict of interest.

利益冲突 作者声明无利益冲突。

1 Zhang S, Roller S, Goyal N, Artetxe M, Chen M, Chen S, Dewan C, Diab M, Li X, Lin XV, Mihaylov T, Ott M, Shleifer S, Shuster K, Simig D, Koura PS, Sridhar A, Wang T, Zettlemoyer L. Opt: open pre-trained transformer language models. arXiv preprint, 2205.01068, 2022

2 Zhang Z, Gu Y, Han X, Chen S, Xiao C, Sun Z, Yao Y, Qi F, Guan J, Ke P, Cai Y, Zeng G, Tan Z, Liu Z, Huang M, Han W, Liu Y, Zhu X, Sun M. AI Open, 2021, 2: 216-224

3 Raffel C, Shazeer N, Roberts A, Lee K, Narang S, Matena M, Zhou Y, Li W, Liu PJ. J Mach Learn Res, 2020, 21: 1-67

4 Brown TB, Mann B, Ryder N, Subbiah M, Kaplan J, Dhariwal P, Neelakantan A, Shyam P, Sastry G, Askell A, Agarwal S, Herbert-Voss A, Krueger G, Henighan T, Child R, Ramesh A, Ziegler DM, Wu J, Winter C, Hesse C, Chen M, Sigler E, Litwin M, Gray S, Chess B, Clark J, Berner C, McCandlish S, Radford A, Sutskever I, Amodei D. Language models are few-shot learners. arXiv preprint, 2005.14165, 2020

5 https://chat.openai.com/

6 Guo T, Guo K, Nan B, Liang Z, Guo Z, Chawla NV, Wiest O, Zhang X. What can large language models do in chemistry? A comprehensive benchmark on eight tasks. arXiv preprint, 2305.18365, 2023

7 Boiko DA, MacKnight R, Kline B, Gomes G. Nature, 2023, 624: 570-578

8 Silver D, Schrittwieser J, Simonyan K, Antonoglou I, Huang A, Guez A, Hubert T, Baker L, Lai M, Bolton A, Chen Y, Lillicrap T, Hui F, Sifre L, van den Driessche G, Graepel T, Hassabis D. Nature, 2017, 550: 354-359

9 Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R, Židek A, Potapenko A, Bridgland A, Meyer C, Kohl SAA, Ballard AJ, Cowie A, Romera-Paredes B, Nikolov S, Jain R, Adler J, Back T, Petersen S, Reiman D, Clancy E, Zielinski M, Steinegger M, Pacholska M, Berghammer T, Bodenstein S, Silver D, Vinyals O, Senior AW, Kavukcuoglu K, Kohli P, Has-sabis D. Nature, 2021, 596: 583-589

10 Corey EJ, Long AK, Rubenstein SD. Science, 1985, 228: 408-418

11 Molga K, Szymkuć S, Grzybowski BA. Acc Chem Res, 2021, 54: 1094-1106

12 Mikulak-Klucznik B, Golębiowska P, Bayly AA, Popik O, Klucznik T, Szymkuć S, Gajewska EP, Dittwald P, Staszewska-Krajewska O, Beker W, Badowski T, Scheidt KA, Molga K, Mlynarski J, Mrksich M, Grzybowski BA. Nature, 2020, 588: 83-88

13 Ucak UV, Ashyrmamatov I, Ko J, Lee J. Nat Commun, 2022, 13: 1186

14 Bragato M, von Rudorff GF, von Lilienfeld OA. Chem Sci, 2020, 11: 11859-11868

15 Ingold CK. Chem Rev, 1934, 15: 225-274

16 Kermack WO, Robinson R. J Chem Soc Trans, 1922, 121: 427-440 17 Hammett LP. J Am Chem Soc, 1937, 59: 96-103

18 Seeman JI. Chem Rev, 1983, 83: 83-134

19 Santiago CB, Guo JY, Sigman MS. Chem Sci, 2018, 9: 2398-2412 20 http://ibond.nankai.edu.cn/

21 Yang Q, Li Y, Yang JD, Liu Y, Zhang L, Luo S, Cheng JP. Angew Chem Int Ed, 2020, 59: 19282-19291

22 Peplow M. Nature, 2014, 512: 20-22

23 Giles J. Nature, 2012, 481: 430-431

24 Degtyarenko K, de Matos P, Ennis M, Hastings J, Zbinden M, McNaught A, Alcantara R, Darsow M, Guedj M, Ashburner M. Nucleic Acids Res, 2007, 36: D344-D350

25 Gaulton A, Bellis LJ, Bento AP, Chambers J, Davies M, Hersey A, Light Y, McGlinchey S, Michalovich D, Al-Lazikani B, Overington JP. Nucleic Acids Res, 2012, 40: D1100-D1107

26 Saito T, Kinugasa S. Synth Engl Ed, 2011, 4: 35-44

27 Gražulis S, Chateigner D, Downs RT, Yokochi AFT, Quirós M, Lutterotti L, Manakova E, Butkus J, Moeck P, Le Bail A. J Appl Crystlogr, 2009, 42: 726-729

28 John Wiley & Sons, Inc. Online spectral database: Quick access to millions of NMR, IR, Raman, UV-vis, and mass spectra. https:// spectrabase.com/

29 Linstrom PJ, Mallard WG. J Chem Eng Data, 2001, 46: 1059-1063

30 Chambers J, Davies M, Gaulton A, Hersey A, Velankar S, Petryszak R, Hastings J, Bellis L, McGlinchey S, Overington JP. J Chemin-form, 2013, 5: 3

31 Gallarati S, van Gerwen P, Laplaza R, Vela S, Fabrizio A, Cormin-boeuf C. Chem Sci, 2022, 13: 13782-13794

32 Irwin JJ, Tang KG, Young J, Dandarchuluun C, Wong BR, Khur-elbaatar M, Moroz YS, Mayfield J, Sayle RA. J Chem Inf Model, 2020, 60: 6065-6073

33 Wang Y, Xiao J, Suzek TO, Zhang J, Wang J, Bryant SH. Nucleic Acids Res, 2009, 37: W623-W633

34 Kim S, Thiessen PA, Bolton EE, Chen J, Fu G, Gindulyte A, Han L, He J, He S, Shoemaker BA, Wang J, Yu B, Zhang J, Bryant SH. Nucleic Acids Res, 2016, 44: D1202-D1213

35 Kearnes SM, Maser MR, Wleklinski M, Kast A, Doyle AG, Dreher SD, Hawkins JM, Jensen KF, Coley CW. J Am Chem Soc, 2021, 143: 18820-18826

36 Ramakrishnan R, Dral PO, Rupp M, von Lilienfeld OA. Sci Data, 2014, 1: 140022

37 Ruddigkeit L, van Deursen R, Blum LC, Reymond JL. J Chem Inf Model, 2012, 52: 2864-2875

38 Weininger D. J Chem Inf Comput Sci, 1988, 28: 31-36

39 Krenn M, Häse F, Nigam AK, Friederich P, Aspuru-Guzik A. Mach Learn-Sci Technol, 2020, 1: 045024

40 Schütt KT, Sauceda HE, Kindermans PJ, Tkatchenko A, Müller KR. J Chem Phys, 2018, 148: 241722

41 Lin TS, Coley CW, Mochigase H, Beech HK, Wang W, Wang Z, Woods E, Craig SL, Johnson JA, Kalow JA, Jensen KF, Olsen BD. ACS Cent Sci, 2019, 5: 1523-1531

42 Guo M, Shou W, Makatura L, Erps T, Foshey M, Matusik W. Adv Sci, 2022, 9: 2101864

43 Bender A, Schneider N, Segler M, Patrick Walters W, Engkvist O, Rodrigues T. Nat Rev Chem, 2022, 6: 428-442

44 Margraf JT. Angew Chem Int Ed, 2023, 62: e202219170

45 Lin Y, Zhang R, Wang D, Cernak T. Science, 2023, 379: 453-457

46 Rodríguez-Pérez R, Bajorath J. J Med Chem, 2020, 63: 8761-8777

47 Rodríguez-Pérez R, Bajorath J. J Comput Aided Mol Des, 2020, 34: 1013-1026

48 Ribeiro MT, Singh S, Guestrin C. Model-agnostic interpretability of machine learning. arXiv preprint, 1606.05386, 2019

49 Zhang WQ, Ge P, Jin WD, Guo J. Radar signal recognition based on TPOT and LIME. In: 2018 37th Chinese Control Conference (CCC). Wuhan, 2018. 4158-4163

50 Lu C, Liu Q, Wang C, Huang Z, Lin P, He L. Molecular property prediction: A multilevel quantum interactions modeling perspective. arXiv preprint, 1906.11081, 2019

51 Liu Y, Yang Q, Li Y, Zhang L, Luo S. Chinese J Org Chem, 2020, 40: 3812-3827

52 Fu Y, Liu L, Li RQ, Liu R, Guo QX. J Am Chem Soc, 2004, 126: 814-822

53 Alongi KS, Shields GC. Theoretical calculations of acid dissociation constants: a review article. In: Ralph W, Ed. Annual Reports in Computational Chemistry. Amsterdam: Elsevier, 2010. 113-138

54 Ho J, Coote ML. WIREs Comput Mol Sci, 2011, 1: 649-660

55 Seybold PG, Shields GC. WIREs Comput Mol Sci, 2015, 5: 290-297

56 Philipp DM, Watson MA, Yu HS, Steinbrecher TB, Bochevarov AD. Int J Quantum Chem, 2018, 118: e25561

57 Wu J, Kang Y, Pan P, Hou T. Drug Discov Today, 2022, 27: 103372

58 Jover J, Bosque R, Sales J. QSAR Comb Sci, 2007, 26: 385-397

59 Harding AP, Wedge DC, Popelier PLA. J Chem Inf Model, 2009, 49: 1914-1924

60 Jover J, Bosque R, Sales J. QSAR Comb Sci, 2008, 27: 1204-1215

61 Chen B, Zhang H, Li M. Neural Comput Applic, 2019, 31: 8297- 8304

62 Zhou T, Jhamb S, Liang X, Sundmacher K, Gani R. Chem Eng Sci, 2018, 183: 95-105

63 Milletti F, Storchi L, Goracci L, Bendels S, Wagner B, Kansy M, Cruciani G. Eur J Medicinal Chem, 2010, 45: 4270-4279

64 Lu Y, Anand S, Shirley W, Gedeck P, Kelley BP, Skolnik S, Rodde S, Nguyen M, Lindvall M, Jia W. J Chem Inf Model, 2019, 59: 4706- 4719

65 Fraczkiewicz R, Lobell M, Göller AH, Krenz U, Schoenneis R, Clark RD, Hillisch A. J Chem Inf Model, 2015, 55: 389-397

66 Roszak R, Beker W, Molga K, Grzybowski BA. J Am Chem Soc, 2019, 141: 17142-17149

67 Mayr F, Wieder M, Wieder O, Langer T. Front Chem, 2022, 10: 866585

68 Feng Y, Liu L, Wang JT, Huang H, Guo QX. J Chem Inf Comput Sci, 2003, 43: 2005-2013

69 Feng Y, Liu L, Wang JT, Zhao SW, Guo QX. J Org Chem, 2004, 69: 3129-3138

70 St. John PC, Guan Y, Kim Y, Etz BD, Kim S, Paton RS. Sci Data, 2020, 7: 244

71 Bosque R, Sales J. J Chem Inf Comput Sci, 2003, 43: 637-642

72 Xue CX, Zhang RS, Liu HX, Yao XJ, Liu MC, Hu ZD, Fan BT. J Chem Inf Comput Sci, 2004, 44: 669-677

73 Nantasenamat C, Isarankura-Na-Ayudhya C, Naenna T, Prachaya-sittikul V. J Mol Graphics Model, 2008, 27: 188-196

74 Xu Q, Xu J. Monatsh Chem, 2016, 148: 645-654

75 Nakajima M, Nemoto T. Sci Rep, 2021, 11: 20207

76 Wen M, Blau SM, Spotte-Smith EWC, Dwaraknath S, Persson KA. Chem Sci, 2021, 12: 1858-1868

77 Gao P, Zhang J, Qiu H, Zhao S. Phys Chem Chem Phys, 2021, 23: 13242-13249

78 Guo S, Jiang J, Ren H, Wang S. J Phys Chem Lett, 2023, 14: 7461- 7468

79 Qu X, Latino DA, Aires-de-Sousa J. J Cheminform, 2013, 5: 34

80 Feng C, Sharman E, Ye S, Luo Y, Jiang J. Sci China Chem, 2019, 62: 1698-1703

81 St. John PC, Guan Y, Kim Y, Kim S, Paton RS. Nat Commun, 2020, 11: 2328

82 Li W, Luan Y, Zhang Q, Aires-de-Sousa J. Mol Inf, 2023, 42: e2200193

83 Unke OT, Meuwly M. J Chem Theor Comput, 2019, 15: 3678-3693

84 Shui Z, Karypis G. Heterogeneous molecular graph neural networks for predicting molecule properties. In: Proceedings - 20th IEEE International Conference on Data Mining. Sorrento, Italy: ICDM, 2020. 492-500

85 Simeon G, De Fabritiis G. TensorNet: cartesian tensor representations for efficient learning of molecular potentials. arXiv preprint, 2306.06482, 2023

86 Fang X, Liu L, Lei J, He D, Zhang S, Zhou J, Wang F, Wu H, Wang H. Nat Mach Intell, 2022, 4: 127-134

87 Zhou G, Gao Z, Ding Q, Zheng H, Xu H, Wei Z, Zhang L, Ke G. Uni-Mol: a universal 3D molecular representation learning framework. ChemRxiv preprint, 2023, doi: 10.26434/chemrxiv-2022- jjm0j-v4

88 Delaney JS. J Chem Inf Comput Sci, 2004, 44: 1000-1005

89 Mobley DL, Guthrie JP. J Comput Aided Mol Des, 2014, 28: 711- 720

90 Hille C, Ringe S, Deimel M, Kunkel C, Acree WE, Reuter K, Oberhofer H. J Chem Phys, 2019, 150: 041710

91 Wu Z, Ramsundar B, Feinberg EN, Gomes J, Geniesse C, Pappu AS, Leswing K, Pande V. Chem Sci, 2018, 9: 513-530

92 Tagade PM, Adiga SP, Park MS, Pandian S, Hariharan KS, Kolake SM. J Phys Chem C, 2018, 122: 11322-11333

93 Ghule S, Dash SR, Bagchi S, Joshi K, Vanka K. ACS Omega, 2022, 7: 11742-11755

94 Grambow CA, Pattanaik L, Green WH. J Phys Chem Lett, 2020, 11: 2992-2997

95 Jorner K, Brinck T, Norrby PO, Buttar D. Chem Sci, 2021, 12: 1163- 1175

96 Houston PL, Nandi A, Bowman JM. J Phys Chem Lett, 2019, 10: 5250-5258

97 Liu Y, Yang Q, Cheng J, Zhang L, Luo S, Cheng JP. Chem-PhysChem, 2023, 24: e202300162

98 Saini V, Sharma A, Nivatia D. Phys Chem Chem Phys, 2022, 24: 1821-1829

99 Orlandi M, Escudero-Casao M, Licini G. J Org Chem, 2021, 86: 3555-3564

100 Hoffmann G, Balcilar M, Tognetti V, Héroux P, Gaüzère B, Adam S, Joubert L. J Comput Chem, 2020, 41: 2124-2136

101 Cuesta SA, Moreno M, López RA, Mora JR, Paz JL, Márquez EA. J Chem Inf Model, 2023, 63: 507-521

102 Boobier S, Liu Y, Sharma K, Hose DRJ, Blacker AJ, Kapur N, Nguyen BN. J Chem Inf Model, 2021, 61: 4890-4899

103 Nie W, Liu D, Li S, Yu H, Fu Y. J Chem Inf Model, 2022, 62: 4319- 4328

104 Käser S, Vazquez-Salazar LI, Meuwly M, Töpfer K. Digital Discov, 2023, 2: 28-58

105 Behler J, Parrinello M. Phys Rev Lett, 2007, 98: 146401

106 Han J, Zhang L, Car R, E W. CiCP, 2018, 23

107 Wang J, Olsson S, Wehmeyer C, Pérez A, Charron NE, de Fabritiis G, Noé F, Clementi C. ACS Cent Sci, 2019, 5: 755-767

108 Bowman JM, Qu C, Conte R, Nandi A, Houston PL, Yu Q. J Chem Theor Comput, 2023, 19: 1-17

109 Kang PL, Shang C, Liu ZP. J Am Chem Soc, 2019, 141: 20525- 20536

110 Ouldridge TE. Nat Comput, 2018, 17: 3-29

111 van Speybroeck V, Gani R, Meier RJ. Chem Soc Rev, 2010, 39: 1764-1779

112 Bell RP. The Proton in Chemistry. Ithaca: Cornell University Press, 1973

113 Stewart R. The Proton: Applications to Organic Chemistry. Orlando: Academic Press, 1985

114 Xue XS, Ji P, Zhou B, Cheng JP. Chem Rev, 2017, 117: 8622-8648

115 Yang JD, Ji P, Xue XS, Cheng JP. J Am Chem Soc, 2018, 140: 8611- 8623

116 Yang JD, Xue J, Cheng JP. Chem Soc Rev, 2019, 48: 2913-2926

117 Cai Z, Liu T, Lin Q, He J, Lei X, Luo F, Huang Y. J Chem Inf Model, 2023, 63: 2936-2947

118 Wei W, Hogues H, Sulea T. J Chem Inf Model, 2023, 63: 5169-5181

119 Luo YR. Comprehensive Handbook of Chemical Bond Energies. Boca Raton: CRC Press, 2007

120 Rupp M, Tkatchenko A, Müller KR, von Lilienfeld OA. Phys Rev Lett, 2012, 108: 058301

121 Ramakrishnan R, Hartmann M, Tapavicza E, von Lilienfeld OA. J Chem Phys, 2015, 143: 084111

122 Blum LC, Reymond JL. J Am Chem Soc, 2009, 131: 8732-8733

123 Di Martino RMC, Maxwell BD, Pirali T. Nat Rev Drug Discov, 2023, 22: 562-584

124 Meanwell NA. J Agric Food Chem, 2023, 71: 18087-18122

125 Bhat V, Welin ER, Guo X, Stoltz BM. Chem Rev, 2017, 117: 4528- 4561

126 Mahboob I, Shafiq I, Shafique S, Akhter P, Amjad US, Hussain M, Park YK. Chem Eng J, 2022, 441: 136063

127 Jorner K, Tomberg A, Bauer C, Sköld C, Norrby PO. Nat Rev Chem, 2021, 5: 240-255

128 Hughes ED. Nature, 1942, 149: 126-130

129 Komp E, Janulaitis N, Valleau S. Phys Chem Chem Phys, 2022, 24: 2692-2705

130 Greaves TL, Schaffarczyk McHale KS, Burkart-Radke RF, Harper JB, Le TC. Phys Chem Chem Phys, 2021, 23: 2742-2752

131 Mayr H, Patz M. Angew Chem Int Ed Engl, 1994, 33: 938-957

132 Mayr H, Ofial AR. SAR QSAR Environ Res, 2015, 26: 619-646

133 Senior AW, Evans R, Jumper J, Kirkpatrick J, Sifre L, Green T, Qin C, Zidek A, Nelson AWR, Bridgland A, Penedones H, Petersen S, Simonyan K, Crossan S, Kohli P, Jones DT, Silver D, Kavukcuoglu K, Hassabis D. Nature, 2020, 577: 706-710

134 Westermayr J, Marquetand P. Mach Learn-Sci Technol, 2020, 1: 043001

135 Unke OT, Chmiela S, Sauceda HE, Gastegger M, Poltavsky I, Schütt KT, Tkatchenko A, Müller KR. Chem Rev, 2021, 121: 10142-10186

136 Noé F, Tkatchenko A, Müller KR, Clementi C. Annu Rev Phys Chem, 2020, 71: 361-390

137 Schaaf LL, Fako E, De S, Schäfer A, Csányi G. npj Comput Mater, 2023, 9: 180

138 Chmiela S, Vassilev-Galindo V, Unke OT, Kabylda A, Sauceda HE, Tkatchenko A, Müller KR. Sci Adv, 2023, 9: eadf0873

139 Shu Y, Varga Z, Jasper A, Espinosa-Garcia J, Corchado JC, Truhlar DG. Comput Phys Commun, 2024, 294: 108937

140 Schlegel HB. J Comput Chem, 2003, 24: 1514-1527

141 Wei GF, Liu ZP. J Chem Theor Comput, 2016, 12: 4698-4706

142 Huang SD, Shang C, Zhang XJ, Liu ZP. Chem Sci, 2017, 8: 6327- 6337

143 Zhang XJ, Shang C, Liu ZP. J Chem Phys, 2017, 147: 152706

144 Zhang XJ, Shang C, Liu ZP. Phys Chem Chem Phys, 2017, 19: 4725-4733

145 Fang YH, Ma SC, Liu ZP. J Phys Chem C, 2019, 123: 19347-19353

146 Corey EJ, Wipke WT. Science, 1969, 166: 178-192

147 Segler MHS, Waller MP. Chem Eur J, 2017, 23: 5966-5971

148 Segler MHS, Preuss M, Waller MP. Nature, 2018, 555: 604-610

149 Kishimoto A, Buesser B, Chen B, Botea A. Depth-first proof-number search with heuristic edge cost and application to chemical synthesis planning. In: Advances in Neural Information Processing Systems 32 (Nips 2019). Vancouver, 2019. 7226-7236

150 Chen B, Li C, Dai H, Song L. Learning retrosynthetic planning with neural guided A* Search. In: Proceedings of the 37th International Conference on Machine Learning. online, 2020. 119: 1608-1616

151 Xie S, Yan R, Han P, Xia Y, Wu L, Guo C, Yang B, Qin T. Retro-graph: retrosynthetic planning with graph search. arXiv preprint, 2206.11477, 2022

152 Kim J, Ahn S, Lee H, Shin J. Self-improved retrosynthetic planning. arXiv preprint, 2106.04880, 2021

153 Yu Y, Wei Y, Kuang K, Huang Z, Yao H, Wu F. GRASP: navigating retrosynthetic planning with goal-driven policy. In: Proceedings of the 36th International Conference on Neural Information Processing Systems. New Orleans, 2022. 10257-10268

154 Liu G, Xue D, Xie S, Xia Y, Tripp A, Maziarz K, Segler M, Qin T, Zhang Z, Liu TY. Retrosynthetic planning with dual value networks. arXiv preprint, 2301.13755, 2023

155 Coley CW, Thomas Iii DA, Lummiss JAM, Jaworski JN, Breen CP, Schultz V, Hart T, Fishman JS, Rogers L, Gao H, Hicklin RW, Plehiers PP, Byington J, Piotti JS, Green WH, Hart AJ, Jamison TF, Jensen KF. Science, 2019, 365: eaax1566

156 Lin Y, Zhang Z, Mahjour B, Wang D, Zhang R, Shim E, McGrath A, Shen Y, Brugger N, Turnbull R, Trice S, Jasty S, Cernak T. Nat Commun, 2021, 12: 7327

157 Wolos A, Koszelewski D, Roszak R, Szymkuć S, Moskal M, Os-taszewski R, Herrera BT, Maier JM, Brezicki G, Samuel J, Lummiss JAM, McQuade DT, Rogers L, Grzybowski BA. Nature, 2022, 604: 668-676

158 Jewett JC, Bertozzi CR. Chem Soc Rev, 2010, 39: 1272-1279

159 Park Y, Kim Y, Chang S. Chem Rev, 2017, 117: 9247-9301

160 Braconi E. Nat Rev Methods Primers, 2023, 3: 74

161 Rinehart NI, Zahrt AF, Henle JJ, Denmark SE. Acc Chem Res, 2021, 54: 2041-2054

162 Żuráński AM, Martinez Alvarado JI, Shields BJ, Doyle AG. Acc Chem Res, 2021, 54: 1856-1865

163 Muratov EN, Bajorath J, Sheridan RP, Tetko IV, Filimonov D, Poroikov V, Oprea TI, Baskin II, Varnek A, Roitberg A, Isayev O, Curtalolo S, Fourches D, Cohen Y, Aspuru-Guzik A, Winkler DA, Agrafiotis D, Cherkasov A, Tropsha A. Chem Soc Rev, 2020, 49: 3525-3564

164 Zahrt AF, Athavale SV, Denmark SE. Chem Rev, 2020, 120: 1620- 1689

165 Oliveira JCA, Frey J, Zhang SQ, Xu LC, Li X, Li SW, Hong X, Ackermann L. Trends Chem, 2022, 4: 863-885

166 Crawford JM, Kingston C, Toste FD, Sigman MS. Acc Chem Res, 2021, 54: 3136-3148

167 Strieth-Kalthoff F, Sandfort F, Segler MHS, Glorius F. Chem Soc Rev, 2020, 49: 6154-6168

168 Yang L, Zhu L, Zhang S, Hong X. Chin J Chem, 2022, 40: 2106- 2117

169 Sandfort F, Strieth-Kalthoff F, Kühnemund M, Beecks C, Glorius F. Chem, 2020, 6: 1379-1390

170 Maley SM, Kwon DH, Rollins N, Stanley JC, Sydora OL, Bischof SM, Ess DH. Chem Sci, 2020, 11: 9665-9674

171 Gallarati S, Fabregat R, Laplaza R, Bhattacharjee S, Wodrich MD, Corminboeuf C. Chem Sci, 2021, 12: 6879-6889

172 Moskal M, Beker W, Szymkuć S, Grzybowski BA. Angew Chem Int Ed, 2021, 60: 15230-15235

173 Li B, Su S, Zhu C, Lin J, Hu X, Su L, Yu Z, Liao K, Chen H. J Cheminform, 2023, 15: 72

174 Tsuji N, Sidorov P, Zhu C, Nagata Y, Gimadiev T, Varnek A, List B. Angew Chem Int Ed, 2023, 62: e202218659

175 Xu Y, Gao Y, Su L, Wu H, Tian H, Zeng M, Xu C, Zhu X, Liao K. Angew Chem Int Ed, 2023, 62

176 Shields BJ, Stevens J, Li J, Parasram M, Damani F, Alvarado JIM, Janey JM, Adams RP, Doyle AG. Nature, 2021, 590: 89-96

177 Gensch T, Smith SR, Colacot TJ, Timsina YN, Xu G, Glasspoole BW, Sigman MS. ACS Catal, 2022, 12: 7773-7780

178 Liles JP, Rouget-Virbel C, Wahlman JLH, Rahimoff R, Crawford JM, Medlin A, O’Connor VS, Li J, Roytman VA, Toste FD, Sigman MS. Chem, 2023, 9: 1518-1537

179 van Dijk L, Haas BC, Lim NK, Clagg K, Dotson JJ, Treacy SM, Piechowicz KA, Roytman VA, Zhang H, Toste FD, Miller SJ, Gosselin F, Sigman MS. J Am Chem Soc, 2023, 145: 20959-20967

180 Ahneman DT, Estrada JG, Lin S, Dreher SD, Doyle AG. Science, 2018, 360: 186-190

181 Qiu J, Xie J, Su S, Gao Y, Meng H, Yang Y, Liao K. Chem, 2022, 8: 3275-3287

182 Saebi M, Nan B, Herr JE, Wahlers J, Guo Z, Zurański AM, Kogej T, Norrby PO, Doyle AG, Chawla NV, Wiest O. Chem Sci, 2023, 14: 4997-5005

183 Schwaller P, Vaucher AC, Laino T, Reymond JL. Mach Learn-Sci Technol, 2021, 2: 015016

184 Strieth-Kalthoff F, Sandfort F, Kühnemund M, Schäfer FR, Kuchen H, Glorius F. Angew Chem Int Ed, 2022, 61: e202204647

185 Gallegos LC, Luchini G, St. John PC, Kim S, Paton RS. Acc Chem Res, 2021, 54: 827-836

186 Singh S, Sunoj RB. Acc Chem Res, 2023, 56: 402-412

187 Zahrt AF, Henle JJ, Rose BT, Wang Y, Darrow WT, Denmark SE. Science, 2019, 363: eaau5631

188 Reid JP, Sigman MS. Nature, 2019, 571: 343-348

189 Zhang X, Chung LW, Wu YD. Acc Chem Res, 2016, 49: 1302-1310

190 Zhang S, Xu L, Li S, Oliveira JCA, Li X, Ackermann L, Hong X. Chem Eur J, 2023, 29: e202202834

191 Xu L, Zhang S, Li X, Tang M, Xie P, Hong X. Angew Chem Int Ed, 2021, 60: 22804-22811

192 Xu LC, Frey J, Hou X, Zhang SQ, Li YY, Oliveira JCA, Li SW, Ackermann L, Hong X. Nat Synth, 2023, 2: 321-330

193 Zhang ZJ, Li SW, Oliveira JCA, Li Y, Chen X, Zhang SQ, Xu LC, Rogge T, Hong X, Ackermann L. Nat Commun, 2023, 14: 3149

194 Beker W, Gajewska EP, Badowski T, Grzybowski BA. Angew Chem Int Ed, 2019, 58: 4515-4519

195 Melville J, Hargis C, Davenport MT, Hamilton RS, Ess DH. J Phys Org Chem, 2022, 35: e4405

196 Pesciullesi G, Schwaller P, Laino T, Reymond JL. Nat Commun, 2020, 11: 4874

197 Li X, Zhang S, Xu L, Hong X. Angew Chem Int Ed, 2020, 59: 13253-13259

198 Guan Y, Coley CW, Wu H, Ranasinghe D, Heid E, Struble TJ, Pattanaik L, Green WH, Jensen KF. Chem Sci, 2021, 12: 2198-2208

199 Boni YT, Cammarota RC, Liao K, Sigman MS, Davies HML. J Am Chem Soc, 2022, 144: 15549-15561

200 Dhawa U, Tian C, Wdowik T, Oliveira JCA, Hao J, Ackermann L. Angew Chem Int Ed, 2020, 59: 13451-13457

201 Caldeweyher E, Elkin M, Gheibi G, Johansson M, Sköld C, Norrby PO, Hartwig JF. J Am Chem Soc, 2023, 145: 17367-17376

202 Beker W, Roszak R, Wolos A, Angello NH, Rathore V, Burke MD, Grzybowski BA. J Am Chem Soc, 2022, 144: 4819-4827

203 Häse F, Roch LM, Kreisbeck C, Aspuru-Guzik A. ACS Cent Sci, 2018, 4: 1134-1145

204 Singh S, Pareek M, Changotra A, Banerjee S, Bhaskararao B, Ba-lamurugan P, Sunoj RB. Proc Natl Acad Sci USA, 2020, 117: 1339- 1345

205 Guo Y, He X, Su Y, Dai Y, Xie M, Yang S, Chen J, Wang K, Zhou D, Wang C. J Am Chem Soc, 2021, 143: 5755-5762

206 Raccuglia P, Elbert KC, Adler PDF, Falk C, Wenny MB, Mollo A, Zeller M, Friedler SA, Schrier J, Norquist AJ. Nature, 2016, 533: 73- 76

207 Bai Y, Wilbraham L, Slater BJ, Zwijnenburg MA, Sprick RS, Cooper AI. J Am Chem Soc, 2019, 141: 9063-9071

208 Doan Tran H, Kim C, Chen L, Chandrasekaran A, Batra R, Venka-tram S, Kamal D, Lightstone JP, Gurnani R, Shetty P, Ramprasad M, Laws J, Shelton M, Ramprasad R. J Appl Phys, 2020, 128: 171104

209 Kim C, Chandrasekaran A, Huan TD, Das D, Ramprasad R. J Phys Chem C, 2018, 122: 17575-17585

210 Pilania G, Wang C, Jiang X, Rajasekaran S, Ramprasad R. Sci Rep, 2013, 3: 2810

211 Wu K, Sukumar N, Lanzillo NA, Wang C, “Rampi” Ramprasad R, Ma R, Baldwin AF, Sotzing G, Breneman C. J Polym Sci Part B-Polym Phys, 2016, 54: 2082-2091

212 Simine L, Allen TC, Rossky PJ. Proc Natl Acad Sci USA, 2020, 117: 13945-13948

213 Wang Y, Xie T, France-Lanord A, Berkley A, Johnson JA, Shao-Horn Y, Grossman JC. Chem Mater, 2020, 32: 4144-4151

214 Aldeghi M, Coley CW. Chem Sci, 2022, 13: 10486-10498

215 Jha A, Chandrasekaran A, Kim C, Ramprasad R. Model Simul Mater Sci Eng, 2019, 27: 024002

216 Sun W, Zheng Y, Yang K, Zhang Q, Shah AA, Wu Z, Sun Y, Feng L, Chen D, Xiao Z, Lu S, Li Y, Sun K. Sci Adv, 2019, 5: eaay4275

217 Xu C, Wang Y, Barati Farimani A. npj Comput Mater, 2023, 9: 64 218 Kuenneth C, Ramprasad R. Nat Commun, 2023, 14: 4099

219 Rahman A, Deshpande P, Radue MS, Odegard GM, Gowtham S, Ghosh S, Spear AD. Compos Sci Tech, 2021, 207: 108627

220 Park J, Shim Y, Lee F, Rammohan A, Goyal S, Shim M, Jeong C, Kim DS. ACS Polym Au, 2022, 2: 213-222

221 Yamada H, Liu C, Wu S, Koyama Y, Ju S, Shiomi J, Morikawa J, Yoshida R. ACS Cent Sci, 2019, 5: 1717-1730

222 Venkatram S, Batra R, Chen L, Kim C, Shelton M, Ramprasad R. J Phys Chem B, 2020, 124: 6046-6054

223 Kuenneth C, Rajan AC, Tran H, Chen L, Kim C, Ramprasad R. Patterns, 2021, 2: 100238

224 Mannodi-Kanakkithodi A, Pilania G, Huan TD, Lookman T, Ram-prasad R. Sci Rep, 2016, 6: 20952

225 Afzal MAF, Haghighatlari M, Ganesh SP, Cheng C, Hachmann J. J Phys Chem C, 2019, 123: 14610-14618

226 Wu S, Kondo Y, Kakimoto M, Yang B, Yamada H, Kuwajima I, Lambard G, Hongo K, Xu Y, Shiomi J, Schick C, Morikawa J, Yoshida R. npj Comput Mater, 2019, 5: 66

227 Kim C, Batra R, Chen L, Tran H, Ramprasad R. Comput Mater Sci, 2021, 186: 110067

228 Zhou T, Wu Z, Chilukoti HK, Müller-Plathe F. J Chem Theor Comput, 2021, 17: 3772-3782

229 Martin TB, Audus DJ. ACS Polym Au, 2023, 3: 239-258

230 Rubens M, Vrijsen JH, Laun J, Junkers T. Angew Chem Int Ed, 2019, 58: 3183-3187

231 Rubens M, Van Herck J, Junkers T. ACS Macro Lett, 2019, 8: 1437- 1441

232 Zhang B, Mathoor A, Junkers T. Angew Chem Int Ed, 2023, 62: e202308838

233 Knox ST, Parkinson SJ, Wilding CYP, Bourne RA, Warren NJ. Polym Chem, 2022, 13: 1576-1585

234 Rizkin BA, Shkolnik AS, Ferraro NJ, Hartman RL. Nat Mach Intell, 2020, 2: 200-209

235 Gu Y, Lin P, Zhou C, Chen M. Sci China Chem, 2021, 64: 1039- 1046

236 Zhao B, Cheng J, Gao J, Haddleton DM, Wilson P. Macro Chem Phys, 2023, 224: 2300039

237 Chapman R, Gormley AJ, Herpoldt KL, Stevens MM. Macromolecules, 2014, 47: 8541-8547

238 Chapman R, Gormley AJ, Stenzel MH, Stevens MM. Angew Chem Int Ed, 2016, 128: 4500-4503

239 Liu Z, Lv Y, An Z. Angew Chem Int Ed, 2017, 56: 13852-13856

240 Li R, Zhang S, Li Q, Qiao GG, An Z. Angew Chem Int Ed, 2022, 61: e202213396

241 Enciso AE, Fu L, Russell AJ, Matyjaszewski K. Angew Chem Int Ed, 2018, 57: 933-936

242 Xu J, Jung K, Atme A, Shanmugam S, Boyer C. J Am Chem Soc, 2014, 136: 5508-5519

243 Gormley AJ, Yeow J, Ng G, Conway Ó, Boyer C, Chapman R. Angew Chem Int Ed, 2018, 57: 1557-1562

244 Zhou Y, Gu C, Zheng L, Shan F, Chen G. Polym Chem, 2022, 13: 989-996

245 Zheng Y, Luo Y, Feng K, Zhang W, Chen G. ACS Macro Lett, 2019, 8: 326-330

246 Jafari VF, Mossayebi Z, Allison-Logan S, Shabani S, Qiao GG. Chem Eur J, 2023, 29: e202301767

247 Lv C, He C, Pan X. Angew Chem Int Ed, 2018, 57: 9430-9433

248 Stubbs C, Congdon T, Davis J, Lester D, Richards SJ, Gibson MI. Macromolecules, 2019, 52: 7603-7612

249 Zhong Y, Zeberl BJ, Wang X, Luo J. Acta Biomater, 2018, 73: 21-37

250 Ladmiral V, Mantovani G, Clarkson GJ, Cauet S, Irwin JL, Had-dleton DM. J Am Chem Soc, 2006, 128: 4823-4830

251 Yan Y, Liu L, Xiong H, Miller JB, Zhou K, Kos P, Huffman KE, Elkassih S, Norman JW, Carstens R, Kim J, Minna JD, Siegwart DJ. Proc Natl Acad Sci USA, 2016, 113: E5702-E5710

252 Upadhya R, Kanagala MJ, Gormley AJ. Macromol Rapid Commun, 2019, 40: 1900528

253 Hughes I, Hunter D. Curr Opin Chem Biol, 2001, 5: 243-247

254 Schmatloch S, Meier MAR, Schubert US. Macromol Rapid Com-mun, 2003, 24: 33-46

255 Zhang S, Wang L, Fu X. Sci Sin Chim, 2023, 53: 3-8

256 Bannigan P, Bao Z, Hickman RJ, Aldeghi M, Häse F, Aspuru-Guzik A, Allen C. Nat Commun, 2023, 14: 35

257 Kumar R, Le N, Tan Z, Brown ME, Jiang S, Reineke TM. ACS Nano, 2020, 14: 17626-17639

258 Ye S, Meftahi N, Lyskov I, Tian T, Whitfield R, Kumar S, Chris-tofferson AJ, Winkler DA, Shih CJ, Russo S, Leroux JC, Bao Y. Chem, 2023, 9: 924-947

259 Fransen KA, Av-Ron SHM, Buchanan TR, Walsh DJ, Rota DT, Van Note L, Olsen BD. Proc Natl Acad Sci USA, 2023, 120: e2220021120

260 Sanders MA, Chittari SS, Sherman N, Foley JR, Knight AS. J Am Chem Soc, 2023, 145: 9686-9692

261 Gormley AJ, Webb MA. Nat Rev Mater, 2021, 6: 642-644

262 Pollice R, dos Passos Gomes G, Aldeghi M, Hickman RJ, Krenn M, Lavigne C, Lindner-D’Addario M, Nigam AK, Ser CT, Yao Z, As-puru-Guzik A. Acc Chem Res, 2021, 54: 849-860

263 Vamathevan J, Clark D, Czodrowski P, Dunham I, Ferran E, Lee G, Li B, Madabhushi A, Shah P, Spitzer M, Zhao S. Nat Rev Drug Discov, 2019, 18: 463-477

264 Reis M, Gusev F, Taylor NG, Chung SH, Verber MD, Lee YZ, Isayev O, Leibfarth FA. J Am Chem Soc, 2021, 143: 17677-17689

265 Kosuri S, Borca CH, Mugnier H, Tamasi M, Patel RA, Perez I, Kumar S, Finkel Z, Schloss R, Cai L, Yarmush ML, Webb MA, Gormley AJ. Adv Healthcare Mater, 2022, 11: 2102101

266 Tamasi MJ, Patel RA, Borca CH, Kosuri S, Mugnier H, Upadhya R, Murthy NS, Webb MA, Gormley AJ. Adv Mater, 2022, 34: 2201809

267 Wu G, Zhou H, Zhang J, Tian ZY, Liu X, Wang S, Coley CW, Lu H. Nat Synth, 2023, 2: 515-526

268 Bordin N, Dallago C, Heinzinger M, Kim S, Littmann M, Rauer C, Steinegger M, Rost B, Orengo C. Trends Biochem Sci, 2023, 48: 345-359

269 Baek M, DiMaio F, Anishchenko I, Dauparas J, Ovchinnikov S, Lee GR, Wang J, Cong Q, Kinch LN, Schaeffer RD, Millán C, Park H, Adams C, Glassman CR, DeGiovanni A, Pereira JH, Rodrigues AV, van Dijk AA, Ebrecht AC, Opperman DJ, Sagmeister T, Buhlheller C, Pavkov-Keller T, Rathinaswamy MK, Dalwadi U, Yip CK, Burke JE, Garcia KC, Grishin NV, Adams PD, Read RJ, Baker D. Science, 2021, 373: 871-876

270 Lin Z, Akin H, Rao R, Hie B, Zhu Z, Lu W, Smetanin N, Verkuil R, Kabeli O, Shmueli Y, dos Santos Costa A, Fazel-Zarandi M, Sercu T, Candido S, Rives A. Science, 2023, 379: 1123-1130

271 Yu T, Cui H, Li JC, Luo Y, Jiang G, Zhao H. Science, 2023, 379: 1358-1363

272 van Kempen M, Kim SS, Tumescheit C, Mirdita M, Lee J, Gilchrist CLM, Söding J, Steinegger M. Nat Biotechnol, 2024, 42: 243-246

273 Alley EC, Khimulya G, Biswas S, AlQuraishi M, Church GM. Nat Methods, 2019, 16: 1315-1322

274 Madani A, Krause B, Greene ER, Subramanian S, Mohr BP, Holton JM, Olmos Jr. JL, Xiong C, Sun ZZ, Socher R, Fraser JS, Naik N. Nat Biotechnol, 2023, 41: 1099-1106

275 Rives A, Meier J, Sercu T, Goyal S, Lin Z, Liu J, Guo D, Ott M, Zitnick CL, Ma J, Fergus R. Proc Natl Acad Sci USA, 2021, 118: e2016239118

276 Jehl P, Manguy J, Shields DC, Higgins DG, Davey NE. Nucleic Acids Res, 2016, 44: W11-W15

277 Elnaggar A, Heinzinger M, Dallago C, Rehawi G, Wang Y, Jones L, Gibbs T, Feher T, Angerer C, Steinegger M, Bhowmik D, Rost B. IEEE Trans Pattern Anal Mach Intell, 2022, 44: 7112-7127

278 Brandes N, Ofer D, Peleg Y, Rappoport N, Linial M. Bioinformatics, 2022, 38: 2102-2110

279 Leaver-Fay A, Tyka M, Lewis SM, Lange OF, Thompson J, Jacak R, Kaufman KW, Renfrew PD, Smith CA, Sheffler W, Davis IW, Cooper S, Treuille A, Mandell DJ, Richter F, Ban Y-EA, Fleishman SJ, Corn JE, Kim DE, Lyskov S, Berrondo M, Mentzer S, Popović Z, Havranek JJ, Karanicolas J, Das R, Meiler J, Kortemme T, Gray JJ, Kuhlman B, Baker D, Bradley P. Meth Enzymol, 2011, 487: 545-574

280 Schymkowitz J, Borg J, Stricher F, Nys R, Rousseau F, Serrano L. Nucleic Acids Res, 2005, 33: W382-W388

281 Liu Y, Zhang L, Wang W, Zhu M, Wang C, Li F, Zhang J, Li H, Chen Q, Liu H. Nat Comput Sci, 2022, 2: 451-462

282 Dauparas J, Anishchenko I, Bennett N, Bai H, Ragotte RJ, Milles LF, Wicky BIM, Courbet A, de Haas RJ, Bethel N, Leung PJY, Huddy TF, Pellock S, Tischer D, Chan F, Koepnick B, Nguyen H, Kang A, Sankaran B, Bera AK, King NP, Baker D. Science, 2022, 378: 49-56

283 Huang B, Xu Y, Hu X, Liu Y, Liao S, Zhang J, Huang C, Hong J, Chen Q, Liu H. Nature, 2022, 602: 523-528

284 Anishchenko I, Pellock SJ, Chidyausiku TM, Ramelot TA, Ovchin-nikov S, Hao J, Bafna K, Norn C, Kang A, Bera AK, DiMaio F, Carter L, Chow CM, Montelione GT, Baker D. Nature, 2021, 600: 547-552

285 Watson JL, Juergens D, Bennett NR, Trippe BL, Yim J, Eisenach HE, Ahern W, Borst AJ, Ragotte RJ, Milles LF, Wicky BIM, Hanikel N, Pellock SJ, Courbet A, Sheffler W, Wang J, Venkatesh P, Sappington I, Torres SV, Lauko A, De Bortoli V, Mathieu E, Ovchinnikov S, Barzilay R, Jaakkola TS, DiMaio F, Baek M, Baker D. Nature, 2023, 620: 1089-1100

286 Ingraham JB, Baranov M, Costello Z, Barber KW, Wang W, Ismail A, Frappier V, Lord DM, Ng-Thow-Hing C, Van Vlack ER, Tie S, Xue V, Cowles SC, Leung A, Rodrigues JV, Morales-Perez CL, Ayoub AM, Green R, Puentes K, Oplinger F, Panwar NV, Ober-meyer F, Root AR, Beam AL, Poelwijk FJ, Grigoryan G. Nature, 2023, 623: 1070-1078

287 Jiang L, Althoff EA, Clemente FR, Doyle L, Rothlisberger D, Zan-ghellini A, Gallaher JL, Betker JL, Tanaka F, Barbas Iii CF, Hilvert D, Houk KN, Stoddard BL, Baker D. Science, 2008, 319: 1387-1391

288 Yeh AHW, Norn C, Kipnis Y, Tischer D, Pellock SJ, Evans D, Ma P, Lee GR, Zhang JZ, Anishchenko I, Coventry B, Cao L, Dauparas J, Halabiya S, DeWitt M, Carter L, Houk KN, Baker D. Nature, 2023, 614: 774-780

289 Yang KK, Wu Z, Arnold FH. Nat Methods, 2019, 16: 687-694

290 Hopf TA, Ingraham JB, Poelwijk FJ, Schärfe CPI, Springer M, Sander C, Marks DS. Nat Biotechnol, 2017, 35: 128-135

291 Shamsi Z, Chan M, Shukla D. J Phys Chem B, 2020, 124: 3845-3854

292 Hsu C, Nisonoff H, Fannjiang C, Listgarten J. Nat Biotechnol, 2022, 40: 1114-1122

293 Biswas S, Khimulya G, Alley EC, Esvelt KM, Church GM. Nat Methods, 2021, 18: 389-396

294 Riesselman AJ, Ingraham JB, Marks DS. Nat Methods, 2018, 15: 816-822

295 Esvelt KM, Carlson JC, Liu DR. Nature, 2011, 472: 499-503

296 Dickinson BC, Leconte AM, Allen B, Esvelt KM, Liu DR. Proc Natl Acad Sci USA, 2013, 110: 9007-9012

297 Thuronyi BW, Koblan LW, Levy JM, Yeh WH, Zheng C, Newby GA, Wilson C, Bhaumik M, Shubina-Oleinik O, Holt JR, Liu DR. Nat Biotechnol, 2019, 37: 1070-1079

298 Yu T, Boob AG, Singh N, Su Y, Zhao H. Cell Syst, 2023, 14: 633- 644

299 Enghiad B, Xue P, Singh N, Boob AG, Shi C, Petrov VA, Liu R, Peri SS, Lane ST, Gaither ED, Zhao H. Nat Commun, 2022, 13: 2697

300 HamediRad M, Chao R, Weisberg S, Lian J, Sinha S, Zhao H. Nat Commun, 2019, 10: 5150

301 Sato K, Akiyama M, Sakakibara Y. Nat Commun, 2021, 12: 941

302 Taubert O, von der Lehr F, Bazarova A, Faber C, Knechtges P, Weiel M, Debus C, Coquelin D, Basermann A, Streit A, Kesselheim S, Götz M, Schug A. Commun Biol, 2023, 6: 913

303 Wang W, Feng C, Han R, Wang Z, Ye L, Du Z, Wei H, Zhang F, Peng Z, Yang J. Nat Commun, 2023, 14: 7266

304 Yao W, Xiong DC, Yang Y, Geng C, Cong Z, Li F, Li BH, Qin X, Wang LN, Xue WY, Yu N, Zhang H, Wu X, Liu M, Ye XS. Nat Synth, 2022, 1: 854-863

305 Fang G, Lin D, Liao K. Chin J Chem, 2023, 41: 1075-1079

306 Alvarado-Urbina G, Sathe GM, Liu WC, Gillen MF, Duck PD, Bender R, Ogilvie KK. Science, 1981, 214: 270-274

307 Caruthers MH. Science, 1985, 230: 281-285

308 Merrifield RB, Stewart JM. Nature, 1965, 207: 522-523

309 Merrifield RB. Science, 1965, 150: 178-185

310 Scaringe SA, Wincott FE, Caruthers MH. J Am Chem Soc, 1998, 120: 11820-11821

311 Plante OJ, Palmacci ER, Seeberger PH. Science, 2001, 291: 1523- 1527

312 Legrand M, Foucard A. J Chem Educ, 1978, 55: 767

313 Deadman BJ, Battilocchio C, Sliwinski E, Ley SV. Green Chem, 2013, 15: 2050-2055

314 Sutherland JD, Tu NP, Nemcek TA, Searle PA, Hochlowski JE, Djuric SW, Pan JY. SLAS Tech, 2014, 19: 176-182

315 Tu NP, Searle PA, Sarris K. SLAS Tech, 2016, 21: 459-469

316 Durrer J, Agrawal P, Ozgul A, Neuhauss SCF, Nama N, Ahmed D. Nat Commun, 2022, 13: 6370

317 Plutschack MB, Pieber B, Gilmore K, Seeberger PH. Chem Rev, 2017, 117: 11796-11893

318 Guidi M, Seeberger PH, Gilmore K. Chem Soc Rev, 2020, 49: 8910- 8932

319 Machida K, Hirose Y, Fuse S, Sugawara T, Takahashi T. Chem Pharm Bull, 2010, 58: 87-93

320 Mijalis AJ, Thomas Iii DA, Simon MD, Adamo A, Beaumont R, Jensen KF, Pentelute BL. Nat Chem Biol, 2017, 13: 464-466

321 Perera D, Tucker JW, Brahmbhatt S, Helal CJ, Chong A, Farrell W, Richardson P, Sach NW. Science, 2018, 359: 429-434

322 Li C, Callahan AJ, Simon MD, Totaro KA, Mijalis AJ, Phadke KS, Zhang G, Hartrampf N, Schissel CK, Zhou M, Zong H, Hanson GJ, Loas A, Pohl NLB, Verhoeven DE, Pentelute BL. Nat Commun, 2021, 12: 4396

323 Li C, Callahan AJ, Phadke KS, Bellaire B, Farquhar CE, Zhang G, Schissel CK, Mijalis AJ, Hartrampf N, Loas A, Verhoeven DE, Pentelute BL. ACS Cent Sci, 2021, 8: 205-213

324 Li C, Zhang G, Mohapatra S, Callahan AJ, Loas A, Gómez-Bom-barelli R, Pentelute BL. Adv Sci, 2022, 9: 2201988

325 Steiner S, Wolf J, Glatzel S, Andreou A, Granda JM, Keenan G, Hinkley T, Aragon-Camarasa G, Kitson PJ, Angelone D, Cronin L. Science, 2019, 363: eaav2211

326 Osipyan A, Shaabani S, Warmerdam R, Shishkina SV, Boltz H, Dömling A. Angew Chem Int Ed, 2020, 59: 12423-12427

327 Sagmeister P, Lebl R, Castillo I, Rehrl J, Kruisz J, Sipek M, Horn M, Sacher S, Cantillo D, Williams JD, Kappe CO. Angew Chem Int Ed, 2021, 60: 8139-8148

328 Ahn GN, Sharma BM, Lahore S, Yim SJ, Vidyacharan S, Kim DP. Commun Chem, 2021, 4: 53

329 Chatterjee S, Guidi M, Seeberger PH, Gilmore K. Nature, 2020, 579: 379-384

330 Nandiwale KY, Hart T, Zahrt AF, Nambiar AMK, Mahesh PT, Mo Y, Nieves-Remacha MJ, Johnson MD, García-Losada P, Mateos C, Rincón JA, Jensen KF. React Chem Eng, 2022, 7: 1315-1327

331 Amara Z, Bellamy JFB, Horvath R, Miller SJ, Beeby A, Burgard A, Rossen K, Poliakoff M, George MW. Nat Chem, 2015, 7: 489-495

332 Bana P, Örkényi R, Lövei K, Lakó Å, Túrós GI, Éles J, Faigl F, Greiner I. BioOrg Medicinal Chem, 2017, 25: 6180-6189

333 Bana P, Szigetvári À, Kóti J, Éles J, Greiner I. React Chem Eng, 2019, 4: 652-657

334 Baranczak A, Tu NP, Marjanovic J, Searle PA, Vasudevan A, Djuric SW. ACS Med Chem Lett, 2017, 8: 461-465

335 Thomson CG, Banks C, Allen M, Barker G, Coxon CR, Lee AL, Vilela F. J Org Chem, 2021, 86: 14079-14094

336 Li J, Ballmer SG, Gillis EP, Fujii S, Schmidt MJ, Palazzolo AME, Lehmann JW, Morehouse GF, Burke MD. Science, 2015, 347: 1221- 1226

337 Li J, Grillo AS, Burke MD. Acc Chem Res, 2015, 48: 2297-2307

338 Blair DJ, Chitti S, Trobe M, Kostyra DM, Haley HMS, Hansen RL, Ballmer SG, Woods TJ, Wang W, Mubayi V, Schmidt MJ, Pipal RW, Morehouse GF, Palazzolo Ray AME, Gray DL, Gill AL, Burke MD. Nature, 2022, 604: 92-97

339 Lehmann JW, Blair DJ, Burke MD. Nat Rev Chem, 2018, 2: 0115

340 Hwang YJ, Coley CW, Abolhasani M, Marzinzik AL, Koch G, Spanka C, Lehmann H, Jensen KF. Chem Commun, 2017, 53: 6649- 6652

341 Baumgartner LM, Dennis JM, White NA, Buchwald SL, Jensen KF. Org Process Res Dev, 2019, 23: 1594-1601

342 Mo Y, Rughoobur G, Nambiar AMK, Zhang K, Jensen KF. Angew Chem Int Ed, 2020, 59: 20890-20894

343 Mo Y, Lu Z, Rughoobur G, Patil P, Gershenfeld N, Akinwande AI, Buchwald SL, Jensen KF. Science, 2020, 368: 1352-1357

344 Sun AC, Steyer DJ, Allen AR, Payne EM, Kennedy RT, Stephenson CRJ. Nat Commun, 2020, 11: 6202

345 van Putten R, Eyke NS, Baumgartner LM, Schultz VL, Filonenko GA, Jensen KF, Pidko EA. ChemSusChem, 2022, 15: e202200333

346 Debrouwer W, Kimpe W, Dangreau R, Huvaere K, Gemoets HPL, Mottaghi M, Kuhn S, van Aken K. Org Process Res Dev, 2020, 24: 2319-2325

347 Adamo A, Beingessner RL, Behnam M, Chen J, Jamison TF, Jensen KF, Monbaliu JCM, Myerson AS, Revalor EM, Snead DR, Stelzer T, Weeranoppanant N, Wong SY, Zhang P. Science, 2016, 352: 61-67

348 Jiang T, Bordi S, McMillan AE, Chen KY, Saito F, Nichols PL, Wanner BM, Bode JW. Chem Sci, 2021, 12: 6977-6982

349 McMillan AE, Wu WWX, Nichols PL, Wanner BM, Bode JW. Chem Sci, 2022, 13: 14292-14299

350 Kitson PJ, Marie G, Francoia JP, Zalesskiy SS, Sigerson RC, Ma-thieson JS, Cronin L. Science, 2018, 359: 314-319

351 Hou W, Bubliauskas A, Kitson PJ, Francoia JP, Powell-Davies H, Gutierrez JMP, Frei P, Manzano JS, Cronin L. ACS Cent Sci, 2021, 7: 212-218

352 Bubliauskas A, Blair DJ, Powell-Davies H, Kitson PJ, Burke MD, Cronin L. Angew Chem Int Ed, 2022, 61: e202116108

353 Manzano JS, Hou W, Zalesskiy SS, Frei P, Wang H, Kitson PJ, Cronin L. Nat Chem, 2022, 14: 1311-1318

354 Skilton RA, Bourne RA, Amara Z, Horvath R, Jin J, Scully MJ, Streng E, Tang SLY, Summers PA, Wang J, Pérez E, Asfaw N, Aydos GLP, Dupont J, Comak G, George MW, Poliakoff M. Nat Chem, 2015, 7: 1-5

355 Fitzpatrick DE, Battilocchio C, Ley SV. Org Process Res Dev, 2016, 20: 386-394

356 Roch LM, Häse F, Kreisbeck C, Tamayo-Mendoza T, Yunker LPE, Hein JE, Aspuru-Guzik A. Sci Robot, 2018, 3: eaat5559

357 Rohrbach S, Šiaučiulis M, Chisholm G, Pirvan PA, Saleeb M, Mehr SHM, Trushina E, Leonov AI, Keenan G, Khan A, Hammer A, Cronin L. Science, 2022, 377: 172-180

358 Li J, Tu Y, Liu R, Lu Y, Zhu X. Adv Sci, 2020, 7: 1901957

359 Burger B, Maffettone PM, Gusev VV, Aitchison CM, Bai Y, Wang X, Li X, Alston BM, Li B, Clowes R, Rankin N, Harris B, Sprick RS, Cooper AI. Nature, 2020, 583: 237-241

360 Zhu Q, Zhang F, Huang Y, Xiao H, Zhao LY, Zhang XC, Song T, Tang XS, Li X, He G, Chong BC, Zhou JY, Zhang YH, Zhang B, Cao JQ, Luo M, Wang S, Ye GL, Zhang WJ, Chen X, Cong S, Zhou D, Li H, Li J, Zou G, Shang WW, Jiang J, Luo Y. Natl Sci Rev, 2022, 9: nwac190

361 Ley SV, Fitzpatrick DE, Ingham RJ, Myers RM. Angew Chem Int Ed, 2015, 54: 3449-3464

362 O’Brien M, Koos P, Browne DL, Ley SV. Org Biomol Chem, 2012, 10: 7031

363 O’Brien AG, Horváth Z, Lévesque F, Lee JW, Seidel-Morgenstern A, Seeberger PH. Angew Chem Int Ed, 2012, 51: 7028-7030

364 Ingham RJ, Battilocchio C, Fitzpatrick DE, Sliwinski E, Hawkins JM, Ley SV. Angew Chem Int Ed, 2015, 54: 144-148

365 Granda JM, Donina L, Dragone V, Long DL, Cronin L. Nature, 2018, 559: 377-381

366 Zeng L, Burton L, Yung K, Shushan B, Kassel DB. J Chromatogr A, 1998, 794: 3-13

367 Koppitz M, Brailsford A, Wenz M. J Comb Chem, 2005, 7: 714-720

368 Guth O, Krewer D, Freudenberg B, Paulitz C, Hauser M, Ilg K. J Comb Chem, 2008, 10: 875-882

369 Xu H, Lin J, Liu Q, Chen Y, Zhang J, Yang Y, Young MC, Xu Y, Zhang D, Mo F. Chem, 2022, 8: 3202-3214

370 Xu H, Lin J, Zhang D, Mo F. Nat Commun, 2023, 14: 3095

371 Herges R. J Chem Inf Comput Sci, 1990, 30: 377-383

372 Caramelli D, Granda JM, Mehr SHM, Cambié D, Henson AB, Cronin L. ACS Cent Sci, 2021, 7: 1821-1830

373 McNally A, Prier CK, MacMillan DWC. Science, 2011, 334: 1114- 1117

374 Troshin K, Hartwig JF. Science, 2017, 357: 175-181

375 Zahrt AF, Mo Y, Nandiwale KY, Shprints R, Heid E, Jensen KF. J Am Chem Soc, 2022, 144: 22599-22610

376 Lan T, An Q. J Am Chem Soc, 2021, 143: 16804-16812

377 Yoon J, Cao Z, Raju RK, Wang Y, Burnley R, Gellman AJ, Barati Farimani A, Ulissi ZW. Mach Learn-Sci Technol, 2021, 2: 045018

378 Aldoseri A, Al-Khalifa KN, Hamouda AM. Appl Sci, 2023, 13: 7082

379 Baum ZJ, Yu X, Ayala PY, Zhao Y, Watkins SP, Zhou Q. J Chem Inf Model, 2021, 61: 3197-3212

380 Tetko IV, Engkvist O, Koch U, Reymond J-, Chen H. Mol Inf, 2016, 35: 615-621

381 Emami FS, Vahid A, Wylie EK, Szymkuć S, Dittwald P, Molga K, Grzybowski BA. Angew Chem Int Ed, 2015, 54: 10797-10801

382 Angello NH, Rathore V, Beker W, Wolos A, Jira ER, Roszak R, Wu TC, Schroeder CM, Aspuru-Guzik A, Grzybowski BA, Burke MD. Science, 2022, 378: 399-405

383 Jia X, Lynch A, Huang Y, Danielson M, Lang’at I, Milder A, Ruby AE, Wang H, Friedler SA, Norquist AJ, Schrier J. Nature, 2019, 573: 251-255

384 Huerta EA, Blaiszik B, Brinson LC, Bouchard KE, Diaz D, Doglioni C, Duarte JM, Emani M, Foster I, Fox G, Harris P, Heinrich L, Jha S, Katz DS, Kindratenko V, Kirkpatrick CR, Lassila-Perini K, Madduri RK, Neubauer MS, Psomopoulos FE, Roy A, Rübel O, Zhao Z, Zhu R. Sci Data, 2023, 10: 487

385 Yano J, Gaffney KJ, Gregoire J, Hung L, Ourmazd A, Schrier J, Sethian JA, Toma FM. Nat Rev Chem, 2022, 6: 357-370

386 Liu J, Hein JE. Nat Synth, 2023, 2: 464-466

387 Chithrananda S, Grand G, Ramsundar B. ChemBERTa: large-scale self-supervised pretraining for molecular property prediction. arXiv preprint, 2010.09885, 2020

388 Ross J, Belgodere B, Chenthamarakshan V, Padhi I, Mroueh Y, Das P. Nat Mach Intell, 2022, 4: 1256-1264

389 Frey NC, Soklaski R, Axelrod S, Samsi S, Gómez-Bombarelli R, Coley CW, Gadepally V. Nat Mach Intell, 2023, 5: 1297-1305

390 Wang X, Jiang S, Hu W, Ye S, Wang T, Wu F, Yang L, Li X, Zhang G, Chen X, Jiang J, Luo Y. J Am Chem Soc, 2022, 144: 16069-16076

391 Bédard AC, Adamo A, Aroh KC, Russell MG, Bedermann AA, Torosian J, Yue B, Jensen KF, Jamison TF. Science, 2018, 361: 1220-1225