Abstract
Biologics drug substance (DS) development is becoming increasingly complex, requiring faster, more robust, and more transparent decision-making across the bioprocess lifecycle. Achieving fully integrated digital bioprocessing demands predictive models that connect traditionally siloed stages—from cell line selection to process optimization and manufacturability assessment. AI and machine learning (ML), when combined with mechanistic understanding, enable digital twins that serve as scalable, predictive representations of bioprocesses and form the foundation of this integration.
At Amgen, we are advancing a portfolio of AI/ML digital twins across DS development to move toward an end-to-end digital ecosystem that accelerates timelines, strengthens process understanding, and enables risk-informed decision-making.
This presentation highlights three industrial case studies illustrating this progression.
In cell line development (CLD), we developed Virtual Clone, a digital twin, that leverages machine learning to predict high-performing clones early in development. The approach integrates structured experimental data with features extracted from microscopic cell images using computer vision. Domain-informed feature engineering, and evaluation of multiple ML algorithms resulted in predictive models that rival expert judgment. Virtual Clone enables earlier prioritization of promising clones, reduces experimental burden, and accelerates CLD.
For cell culture process development, we implemented AI/ML digital twins that capture complex, nonlinear relationships among process conditions, media composition, and DS performance attributes. These models enable rapid in silico exploration of operating strategies, supporting process optimization, scale-up readiness, and risk-based decision-making for commercialization while reducing reliance on laboratory experimentation.
To address manufacturability, we developed a hybrid mechanistic–ML digital twin for DS viscosity prediction. By combining first-principles rheological models with ML, the framework preserves interpretability while capturing nonlinear behavior. The model supports viscosity prediction, sensitivity analysis, and cross-process comparison to inform formulation and manufacturing strategy.
Together, these case studies demonstrate how AI/ML digital twins can be embedded across DS development stages, laying the groundwork for fully integrated digital bioprocessing. We also describe a risk-informed model qualification framework that defines context of use, assesses model risk, and establishes credibility through verification, validation, uncertainty quantification, and lifecycle management to enable GMP-aligned in silico decision-making. Finally, we discuss emerging applications of agentic and generative AI to enhance model interaction, workflow integration, and decision support within a connected digital ecosystem.
