Search

Digital twins promise to reshape pharmaceutical development and manufacturing by creating virtual replicas that mirror physical systems over time. Their value lies in linking process knowledge, real-time data, predictive modelling, and decision support into a single operational framework. In pharmaceutical contexts, this promise is especially relevant because product quality is tightly coupled to process history, material attributes, and patient-facing performance. The core problem is that digital twin development in pharma remains fragmented across manufacturing, drug delivery, regulatory modelling, and digital transformation literatures. Manufacturing studies often focus on process control, PAT, and continuous production, whereas drug delivery studies emphasise physiological prediction, formulation performance, and patient-specific behaviour. These areas share mechanistic foundations, but they are rarely treated as parts of a unified pharmaceutical digital twin ecosystem. This conceptual review analyses digital twin logic across pharmaceutical manufacturing and drug delivery systems. It focuses on how mechanistic models, hybrid modelling, and real-time data infrastructures can be combined to support quality, performance prediction, and regulatory decision-making. The central argument is that digital twins must become not only predictive but also explainable and governable. The synthesis defines the architecture of pharmaceutical digital twins, catalogues their manufacturing and drug delivery applications, and identifies unresolved challenges in model coupling, parameter identifiability, uncertainty handling, and regulatory credibility. It also maps technological, organisational, economic, and regulatory barriers that prevent promising models from becoming routine industrial tools. Five tables summarise the conceptual architecture, application domains, model architectures, integration problems, and implementation barriers. Realising the full potential of pharmaceutical digital twins will require mechanistic rigour, explainable analytics, high-quality data connectivity, and early alignment with regulatory expectations. The future of the field will depend less on isolated demonstrations and more on reusable validation strategies, transparent model governance, and cross-sector collaboration. Digital twins should therefore be understood as evolving regulatory-scientific infrastructures rather than as standalone computational artefacts.

Artificial intelligence is rapidly being adopted to optimise pharmaceutical formulation and manufacturing processes, yet its inherent opacity poses a fundamental challenge to regulatory frameworks built on transparency, scientific justification, and mechanistic understanding. In pharmaceutical development, AI systems may influence formulation selection, process parameter optimisation, release modelling, stability prediction, and scale-up strategy. These decisions can directly or indirectly affect product quality and patient safety. The regulatory issue is therefore not whether AI can improve development efficiency, but whether its outputs can be explained, justified, verified, and governed. There is currently no clear regulatory consensus on what constitutes sufficient explainability for AI models used in pharmaceutical formulation and process development. Existing expectations for pharmaceutical development presume that sponsors can describe the relationship between material attributes, process parameters, critical quality attributes, and clinical performance. Many AI models, particularly neural networks, ensemble methods, and adaptive systems, challenge this assumption because their internal decision logic may not be readily interpretable. This uncertainty creates difficulty for industry, regulators, and quality units seeking to evaluate whether AI-supported decisions are scientifically sound. This article identifies the regulatory gaps and risk domains associated with non-explainable AI in pharmaceutical development. It examines how AI is being used across formulation design, process development, manufacturing optimisation, and data-rich pharmaceutical quality systems. It then defines explainability and evidence requirements appropriate for different levels of regulatory risk. The central argument is that explainability should be treated as a regulatory quality attribute of AI systems, not merely as a technical preference. The proposed framework categorises AI applications according to their intended use, model complexity, degree of influence on quality decisions, and potential impact on patient safety. It links these categories to evidence requirements, including training data documentation, performance validation, uncertainty assessment, interpretability justification, human oversight, and lifecycle change controls. Four tables present the AI application landscape, regulatory gap analysis, explainability requirements, and the proposed framework. Together, these elements provide a structured basis for regulatory dialogue and future guidance development. A structured, risk-based approach to regulatory explainability can enable responsible adoption of AI while protecting the integrity of the pharmaceutical regulatory system. Low-risk AI tools may require documented performance and traceability, whereas high-risk tools that influence critical quality decisions require stronger interpretability, validation, and governance. The proposed approach does not require all models to be fully transparent, but it does require that the explanation provided be proportionate to the decision being supported. Regulatory explainability is therefore presented as a necessary bridge between innovation, quality assurance, and patient protection.