At present, with its gain in popularity and widespread utilization, almost all scientific fields are being influenced by the revolution in artificial intelligence (AI). Consequently, many industries are being transformed, signaling a new era. In the past few years, the progress of AI in academic and industrial settings has been notably propelled by the implementation of machine learning (ML) techniques, which is a subset of AI. In addition, the subset of ML known as deep learning (DL), particularly over the past decade, has had a particularly significant impact on various areas. (1) Its influence extends across diverse scientific fields including medicine, aerospace, climate, and many other disciplines. Compared to ML, multiple layers of distributed representations originating from the inputs are uncovered by DL, where more abstract concepts are signified by deeper or higher layers. (2) Data are well interpreted by these DL representations and lead to a substantial improvement in data analysis across a wide range of research fields. Moreover, the advancement of innovative AI/ML algorithms and architectures is expedited significantly by growing computational power, the availability of larger data sets, as well as improved data quality. Over the past decades, there has been a massive development and utilization of modern AI/ML techniques in the pharmaceutical industry, particularly in areas like early drug discovery, as well as drug design and development. Though the pharmaceutical industry continues to extensively employ descriptor-based quantitative structure–activity relationship (QSAR) methods rooted in traditional ML techniques, (3) modern ML approaches like “generative AI” utilizing deep neural networks and representation learning, (4) such as graph-based neural networks are emerging techniques for extracting information from molecular images (2D or 3D chemical structures), and appear to exhibit an enhanced predictivity. In addition, de novo drug design for generating innovative compounds with desirable properties but without a preliminary template has emerged as a potential direction in the pharmaceutical industry. With the emergence and flourishing of deep reinforcement learning techniques, de novo drug design has offered promising opportunities for developing new and effective drugs. (5) For instance, convolutional neural networks (CNN), recurrent neural networks (RNN), generative adversarial networks, and variational autoencoders can be applied widely in pharmaceutical development. Particularly the deep neural network model can solve some of the current challenges, thereby helping to improve the performance of pharmacokinetic─pharmacodynamic (PKPD), physiologically based pharmacokinetic (PBPK) modeling, and mechanistic quantitative systems pharmacology (QSP) models to support early drug discovery and development, (6,7) as well as the prediction of drug compound in the process of absorption, distribution, metabolism, and excretion (ADME). (8) While the QSAR method continues to be utilized, it is now expanding into the realm of computational toxicology. It is regarded as a promising approach to align with the principles of the 3Rs concept which are the replacement, reduction, and refinement of animal testing. (3) Several emerging methods are increasingly gaining attention in the pharmaceutical industry focusing on advanced deep neural networks like CNN and RNN. Additionally, methods involving representation and transfer learning, the application of graph neural networks, and image processing are being employed to predict and elucidate the toxicity of molecules based on their structures or assay responses. Methods for text mining that rely on natural language processing (NLP), the incorporation of chemical structure with omic data sets, and the application of systems biology and pharmacology techniques to anticipate adverse drug effects are also generating significant interest in the toxicology area. (3) Today, dosage form design and drug formulation optimization are critical aspects of drug development. The use of AI/ML allows the pharmaceutical industry to create formulations that ensure the right dose is delivered at the right time, optimizing drug effectiveness and patient compliance. (5) Through large data analysis, AI can boost a drug’s efficacy and bioavailability by enhancing the formulation. (9) More importantly, various algorithms and models, such as multiple linear regression (MLR), random forest (RF), light gradient boosting machine (lightGBM), and neural networks (NN), can be instrumental in predicting drug properties such as solubility, stability, drug release rate, and release kinetics. In addition, personalized medicine is revolutionizing healthcare in the pharmaceutical industry, and AI/ML is driving this shift in dosage form design and delivery. (10) By evaluating the genetic, physiological, and clinical data of the patient, AI/ML models can customize drug formulations and delivery approaches to match the specific requirements of the patient. In pharmaceutical processes, AI/ML models are integrated into cutting-edge manufacturing technologies, including continuous manufacturing, (11) to aid in the advancement of Industry 4.0, which encompasses elements like the Internet of Things (IoT) and Digital Twins (DTs). This fourth industrial revolution is marked by integrated, autonomous, and self-organizing production systems, and it is continuously advancing in pharmaceutical development. (12) Furthermore, AI/ML will continue to play a vital role in optimizing manufacturing processes and scale-up, overseeing processes, identifying faults, and monitoring trends in pharmaceutical development. For clinical trials in pharmaceutical development, AI/ML models can optimize the trial design and selection of patients. For instance, the use of AI/ML can aid in recognizing patients through specific disease characteristics and clinical factors. Additionally, AI can be utilized in trial registries to assess how clinical trial eligibility criteria affect the likelihood of trial termination. (13) Based on the clinical data, the reinforcement learning model can assist in making decisions regarding the sequencing and timing of chemotherapy regimens in a clinical trial for nonsmall-cell lung cancer. (14) Moreover, the random forest model can quickly predict missing values, thereby minimizing data loss and enhancing the analysis of trial data. (15) In recent years, the growing size of data ranging over 100 million or even a billion, (16−18) and a large number of drug molecules, have imposed a challenging demand for developing computational resources, hardware, and advanced algorithms. The use of AI/ML models in pharmaceutical development can be integrated and powered through cloud computing platforms like Amazon Web Services (AWS). This integration empowers the generation of novel scientific insights and the development of advanced AI/ML models to assess the risk of atherosclerosis development. (19,20) Moreover, there is a consistent release of robust hardware designed to enhance the speed of pharmaceutical drug development. This hardware integrates AI/ML, data analytics, simulation, and visualization to facilitate cross-disciplinary workflows in drug design and development such as NVIDIA’s graphic processing unit (GPU), data processing unit (DPU), and Google Cloud’s tensor processing unit (TPU). Furthermore, pharmaceutical development is a domain where the capabilities of ChatGPT, a large language model (LLM) developed by OpenAI, can be effectively used to advance drug discovery and research in previously unattainable ways. For instance, ChatGPT exhibits the capability to predict the PKPD profile and toxicity features of a specific compound, offering essential insights for drug development and protein drug design. (21) Recently, as quantum computing technology has progressed significantly, accelerated pharmaceutical development is poised to gain additional advantages from this advancement in quantum computing technology. (22) The recent progress in AI/ML models for quantum computing has opened up numerous possibilities to broaden the potential applications of machine learning in areas such as drug discovery, toxicology, and the design of dosage forms. (23−26) The use of these innovative technologies and advanced machine learning algorithms has sparked significant enthusiasm and expectation regarding the potential of AI to transform the pharmaceutical industry. AI/ML stands at the forefront of the transformation in pharmaceutical development, serving as powerful tools that reshape the processes of discovery, development, and innovation in the pharmaceutical industry. The successful incorporation of AI/ML holds the promise of expediting research and boosting efficiency in a novel era for pharmaceutical development. Their capacity to analyze intricate biological data, forecast molecular interactions, and streamline decision making carries the potential to hasten pharmaceutical development, cut down costs, and enhance patient outcomes. This article references 26 other publications. This article has not yet been cited by other publications. This article references 26 other publications.

中文翻译：

---

人工智能在药物开发中可能走向何方？

目前，随着人工智能的普及和广泛应​​用，几乎所有科学领域都受到人工智能（AI）革命的影响。因此，许多行业正在转型，标志着一个新时代的到来。在过去的几年里，人工智能在学术和工业领域的进步主要受到机器学习（ML）技术（人工智能的一个子集）的实施的推动。此外，机器学习的一个子集，即深度学习 (DL)，特别是在过去十年中，对各个领域产生了特别重大的影响。(1) 其影响力遍及医学、航空航天、气候等众多学科领域。与 ML 相比，DL 揭示了源自输入的多层分布式表示，其中更抽象的概念由更深或更高的层表示。(2) 这些深度学习表示可以很好地解释数据，并导致广泛研究领域的数据分析得到显着改进。此外，通过计算能力的增强、更大数据集的可用性以及数据质量的提高，创新型人工智能/机器学习算法和架构的进步显着加快。在过去的几十年里，现代人工智能/机器学习技术在制药行业得到了大规模的开发和利用，特别是在早期药物发现以及药物设计和开发等领域。尽管制药行业继续广泛采用植根于传统机器学习技术的基于描述符的定量构效关系 (QSAR) 方法，(3) 现代机器学习方法，例如利用深度神经网络和表示学习的“生成式人工智能”，(4) 例如基于图的神经网络是从分子图像（2D 或 3D 化学结构）中提取信息的新兴技术，并且似乎表现出增强的预测能力。此外，用于生成具有所需特性但没有初步模板的创新化合物的从头药物设计已成为制药行业的潜在方向。随着深度强化学习技术的出现和蓬勃发展，从头药物设计为开发新的有效药物提供了有希望的机会。（5）例如，卷积神经网络（CNN）、循环神经网络（RNN）、生成对抗网络和变分自动编码器可以广泛应用于药物开发。特别是深度神经网络模型可以解决当前的一些挑战，从而有助于提高药代动力学药效学（PKPD）、基于生理学的药代动力学（PBPK）模型和机械定量系统药理学（QSP）模型的性能，以支持早期药物发现和开发，(6,7)以及药物化合物在吸收、分布、代谢和排泄过程中的预测(ADME)。(8) 虽然 QSAR 方法继续被使用，但它现在正在扩展到计算毒理学领域。它被认为是一种符合 3R 概念原则（即动物试验的替代、减少和改进）的有前途的方法。(3) 一些新兴方法越来越受到制药行业的关注，重点关注 CNN 和 RNN 等先进深度神经网络。此外，涉及表示和迁移学习、图神经网络应用和图像处理的方法被用来根据分子的结构或测定响应来预测和阐明分子的毒性。依赖于自然语言处理（NLP）的文本挖掘方法、化学结构与组学数据集的结合，以及应用系统生物学和药理学技术来预测药物不良反应也引起了毒理学领域的极大兴趣。(3) 如今，剂型设计和药物配方优化是药物开发的关键方面。AI/ML 的使用使制药行业能够创建配方，确保在正确的时间提供正确的剂量，从而优化药物有效性和患者依从性。（5）通过大数据分析，人工智能可以通过优化配方来提高药物的疗效和生物利用度。(9) 更重要的是，各种算法和模型，例如多元线性回归（MLR）、随机森林（RF）、光梯度增强机（lightGBM）和神经网络（NN），可以有助于预测药物特性，例如溶解度、稳定性、药物释放速率和释放动力学。此外，个性化医疗正在彻底改变制药行业的医疗保健，而人工智能/机器学习正在推动剂型设计和交付的转变。(10)通过评估患者的遗传、生理和临床数据，AI/ML模型可以定制药物配方和给药方式，以满足患者的具体要求。在制药流程中，人工智能/机器学习模型被集成到尖端制造技术中，包括连续制造，(11) 以帮助推进工业 4.0，其中包括物联网 (IoT) 和数字孪生 (DT) 等元素。第四次工业革命的特点是生产系统集成化、自主化、自组织化，并在医药开发方面不断推进。(12) 此外，人工智能/机器学习将继续在优化制造工艺和扩大规模、监督流程、识别故障和监控药品开发趋势方面发挥至关重要的作用。对于药物开发的临床试验，AI/ML模型可以优化试验设计和患者选择。例如，人工智能/机器学习的使用可以帮助通过特定的疾病特征和临床因素来识别患者。此外，人工智能可用于试验注册，以评估临床试验资格标准如何影响试验终止的可能性。(13)基于临床数据，强化学习模型可以协助非小细胞肺癌临床试验中化疗方案的排序和时机决策。(14)此外，随机森林模型可以快速预测缺失值，从而最大限度地减少数据丢失并增强试验数据的分析。(15) 近年来，数据规模不断增长，超过 1 亿甚至十亿 (16−18) 和大量药物分子，对开发计算资源、硬件和先进算法提出了挑战性的需求。药物开发中 AI/ML 模型的使用可以通过 Amazon Web Services (AWS) 等云计算平台进行集成和支持。这种集成有助于产生新颖的科学见解和开发先进的人工智能/机器学习模型，以评估动脉粥样硬化发展的风险。(19,20) 此外，不断发布旨在提高药物开发速度的强大硬件。该硬件集成了 AI/ML、数据分析、模拟和可视化，以促进药物设计和开发中的跨学科工作流程，例如 NVIDIA 的图形处理单元 (GPU)、数据处理单元 (DPU) 和 Google Cloud 的张量处理单元 (TPU) ）。此外，在药物开发领域，ChatGPT（OpenAI 开发的大型语言模型 (LLM)）的功能可以有效地以以前无法实现的方式推进药物发现和研究。例如，ChatGPT 具有预测特定化合物的 PKPD 特征和毒性特征的能力，为药物开发和蛋白质药物设计提供重要见解。(21) 最近，随着量子计算技术的显着进步，加速的药物开发有望从量子计算技术的进步中获得额外的优势。(22) 量子计算人工智能/机器学习模型的最新进展为拓宽机器学习在药物发现、毒理学和剂型设计等领域的潜在应用开辟了多种可能性。(23−26) 这些创新技术和先进机器学习算法的使用引发了人们对人工智能改变制药行业潜力的极大热情和期望。AI/ML 站在药物开发转型的最前沿，作为重塑制药行业发现、开发和创新流程的强大工具。人工智能/机器学习的成功结合有望在药物开发的新时代加快研究速度并提高效率。他们分析复杂生物数据、预测分子相互作用的能力，简化决策有可能加速药物开发、降低成本并改善患者治疗效果。本文引用了其他 26 篇出版物。这篇文章尚未被其他出版物引用。本文引用了其他 26 篇出版物。