79 TWENTYFOURSEVENBIOPHARMA Issue 3 / October 2025 PHARMATECH ASSOCIATES analytics, one that requires disciplined foundations to ensure scientific integrity, regulatory acceptance and demonstrable business impact, built on a capable framework. Data as a product Viewing data as a product rather than a byproduct of drug development is at the core of preparing an organisation to leverage modelling. Data should be managed, curated and enhanced continuously to deliver ongoing value, much like a digital product, rather than viewed as an artifact of experimental and operational activities across siloed and fragmented systems. Robust data governance is essential and should include validation, traceability, version control, and lifecycle oversight. The complexity and organisational inertia associated with implementing these frameworks within an organisation is significant. Many pharmaceutical and biotech companies face the challenges of fragmented legacy systems, talent shortages or the need to shift workflows and culture to support rigorous, end-to-end data management. Model credibility There has been a lot written about AI model error due to overfitting and ‘hallucinations. The FDA has defined model credibility as the demonstrated trustworthiness of a model’s outputs, substantiated by systematic evidence, for its intended regulatory application. Model credibility is always assessed in relation to its ‘context of use’ (COU), meaning the specific role and decision it is intended to support, such as in non-clinical, clinical, post-marketing or manufacturing applications. Robust model credibility and the mitigation of overfitting are at the core of successful in silico predictive modelling. As advanced analytics and in silico tools become increasingly central to drug discovery, process optimisation and regulatory submission, organisations must anchor these efforts in a foundation of technical, procedural and crossdisciplinary best practices. As regulatory expectations evolve toward model transparency, validation and risk alignment, organisations must also consider the type of modelling framework best suited to their needs. While empirical and AI-driven models draw strength from large, diverse datasets, they are hampered by lack of data and the need for proactive data hygiene and governance. Many questions in drug development require deeper mechanistic understanding, rooted in physical laws, biochemical pathway and systems biology. This is where first-principle and mechanistic modelling offer clear advantages. By embedding established scientific relationships, such as mass transfer, kinetics and molecular interactions directly into the model structure, mechanistic approaches provide interpretable, causally-linked predictions that extend beyond the range of available training data. Unlike purely empirical models, they allow simulation of system behaviour under novel scenarios, supporting process scale-up, optimisation and regulatory submissions with a higher degree of confidence and scientific justification. Mechanistic models, grounded in physical laws, offer unmatched interpretability and regulatory defensibility. Firstprinciples approaches align with ICH Q8(R2) and FDA guidance on continuous manufacturing and are increasingly complemented by AI-driven hybrid models to reduce experimental burden while maintaining predictive power (Zhang et al., 2024)10. Implementation framework Successful deployment of in silico modelling requires deliberate organisational strategy addressing four critical foundations. Organisations must first shift from viewing data as an experimental byproduct to treating it as a strategic asset, implementing robust validation, traceability, version control and lifecycle oversight. For this, many sponsors find value in partnering with specialised providers to access turnkey expertise and validated platforms while building internal capabilities over time. Model credibility and validation represent the second pillar, where the FDA’s context-ofuse framework demands systematic evidence demonstrating model trustworthiness for intended applications. Proper data splitting, cross-validation and uncertainty quantification protect against overfitting while ensuring regulatory defensibility. Model approach and structure selection forms the third foundation, where first-principles and mechanistic approaches offer advantages in interpretability and regulatory acceptance by embedding established scientific relationships without being constrained by the quantity of data. Hybrid models combining mechanistic understanding with AI-driven pattern recognition provide optimal balance of predictive power and scientific justification. Finally, organisational change management requires workflow modification, talent development and cultural adaptation toward data-driven decisionmaking. Early engagement with regulators, clearly defined contexts of use and transparent documentation addressing uncertainty are essential for streamlined review.
RkJQdWJsaXNoZXIy MjY2OTA4MA==