Autonomous AI Assisted Salt Selection for Stability and Performance

该项目将提供一个人工智能驱动的决策系统，用于制药原料药的预测性盐选择。此外，该项目将探索药物材料科学(结构-工艺-性能-性能)的关系，这些关系将指导我们的结晶分类系统开发(CCS)。这项工作将建立我们的行业需求导向计划，与CMAC的未来实验室愿景保持一致，通过将离线测量、协作机器人、分析和决策整合到一个自主的智能开发平台或数据工厂来加快产品和工艺设计。背景在英国，2018年制药业的商品营业额为490亿GB，零售总收入为50亿GB(英国国家统计局，2018年)。此外，在2018年，有人发现，到2035年，人工智能(AI)应用程序可以为英国经济增加6300亿GB的额外收入(Deep Knowledge Analytics，2018，AI in the UK)。这受到了人工智能在广泛领域的进步的刺激，例如材料发现(Schmidt等人，2019年，自然计算材料)，并在政府战略中引起了回应，例如苏格兰的人工智能战略公开呼吁。最近，寻求在制造中建立以人为本、弹性和可持续性的行业5.0原则的出现：这些目标在药品制造领域特别重要。制药行业面临着越来越多的要求，要求更快、更经济和更可持续地提供更多种类的产品。为了实现这一目标，该部门寻求更好地利用网络物理生产系统，特别关注数字双胞胎以及如何使用它们在开发过程中模拟制造过程，从而减少对材料的需求，以在运营前验证过程(Kemppainen，2017，欧洲药物评论)。作为一种提纯和形成与药物相关的多晶型的方法(Lee，2014，《亚洲药学杂志》；Kesisoglou，等，2008，The AAPS Journal)，活性药物成分(API)的结晶是影响下游工艺的制药制造中的关键步骤。确定与工业相关的原料药结晶方法可能是资源密集型的，因为候选结晶过程受到原料药的工业相关溶解度、下游加工实用性和监管机构确定的关键质量属性(CQA)的限制和评估(Chen等人，2011年，Crystal Growth&Design；Brown等人，2018年，分子系统设计与工程)。关键的是，由于市场上约40%的原料药实际上是不溶的，盐(和/或共晶)选择是原料药形式选择过程中的关键步骤，这可以使水的溶解度增加1000倍。然而，适当的反离子选择在很大程度上仍然是反复试验，生成的形式的稳定性和性能(歧化)可能受到盐的晶体结构、配方产品中赋形剂对微环境pH和歧化的影响、盐和赋形剂的缓冲能力的影响。CMAC现有的数据工厂由EPSRC Hub和英国RPIF Net Zero Pilot支持资助，使用高通量小规模结晶实验结合机器学习来减少与探索原料药结晶设计空间相关的时间和材料成本。此外，该项目正在开发一个独特的结晶参数数据库，该数据库将推动进一步的基础和应用研究，朝着开发结晶分类系统(CCS)的方向发展，该系统将自动识别任何给定原料药的与工业相关的实验条件，以实现所需的性能、稳定性或可制造性目标。

This project will deliver an AI-driven decision-making system for predictive salt selection for pharmaceutical APIs. In addition, the project will explore the pharmaceutical materials science (structure-process-property-performance) relationships that inform our Crystallisation Classification System development (CCS). The work will build our industry demand led programme aligned to CMAC's Lab of the Future vision for to accelerate product and process design by incorporating off-line measurement, collaborative robotics, analysis and decisions into an autonomous smart development platform or DataFactory.BackgroundIn the UK the pharmaceutical industry in 2018 was responsible for a £49bn turnover of goods, with a gross revenue from retail sales of £5bn (Office for National Statistics (UK), 2018). Also, in 2018, it was identified that artificial intelligence (AI) applications could add an additional £630bn to the UK economy by 2035 (Deep Knowledge Analytics, 2018, AI in the UK). This has been stimulated by advances in AI in a wide range of sectors, e.g. materials discovery (Schmidt, et al., 2019, Nature Computational Materials), and has provoked responses in government strategy, e.g. Scotland's AI Strategy Open Call. Most recently the emergence of Industry 5.0 principles that seek to build in human centricity, resilience and sustainability to manufacturing: these goals are especially important in the medicines manufacturing area.The pharmaceutical industry is facing increasing demands to deliver a higher variety of products more quickly, economically and sustainably. To achieve this, the sector seeks to make better use of cyber-physical production systems with a specific focus on digital twins and how they can be used to simulate manufacturing processes during development and hence reduce material requirements to validate processes before operation (Kemppainen, 2017, European Pharmaceutical Review). As a method of purification and forming the pharmaceutically relevant polymorph (Lee, 2014, Asian Journal of Pharmaceutical Sciences; Kesisoglou, et al., 2008, The AAPS Journal), crystallisation of the active pharmaceutical ingredient (API) is a key step in pharmaceutical manufacturing that impacts downstream processes. Determining an industrial-relevant approach for API crystallisation can be resource-intensive as a candidate crystallisation process is constrained by and assessed against industrial relevant solubilities, downstream processing practicalities, and regulator-determined Critical Quality Attributes (CQA) of the API (Chen, et al.,2011, Crystal growth & design; Brown et al., 2018, Molecular Systems Design & Engineering). Critically as ca. 40% of marketed APIs are practically insoluble, salt (and/or co-crystal) selection is a key step in the API form selection process which can provide up to 1000 fold increase in aqueous solubility. Yet suitable counterion selection remains largely trial and error and the resultant form stability and performance (disproportionation) can be impacted by salt crystal structure; impact of excipients in the formulated product on microenvironmental pH and disproportionation; buffer capacity of salts and excipients. The existing DataFactory at CMAC funded via EPSRC Hub and UK RPIF Net Zero Pilot support, uses high throughput small-scale crystallisation experiments coupled with machine learning to reduce the time and material costs associated with exploring the crystallisation design space for APIs. In addition the project is developing a unique Crystallisation Parameter Database that will fuel further basic and applied research moving towards the development of a Crystallisation Classification System (CCS) that will autonomously identifies industrially-relevant experimental conditions for any given API to achieve required performance, stability or manufacturability objectives.