Deep-Learning- versus Hypothesis-Driven Modeling in Model-Informed Drug Development: A PK/PD Case Study

Model-informed drug development (MIDD) has traditionally been guided by hypothesis-driven modeling, in which models are constructed based on established biological mechanisms, physicochemical principles, and experimental hypotheses. More recently, the rise of artificial intelligence has enabled data-driven modeling approaches that often bypass explicit mechanistic assumptions. The potential synergy between these two paradigms is reshaping strategies for implementing MIDD. This study aims to compare deep-learning-driven and hypothesis-driven approaches, highlighting their respective strengths, limitations, and opportunities for integration. A case study is presented with the re-analysis of warfarin PK and PK/PD using both modeling paradigms. The comparative PK analysis indicated that the deep-learning model achieved predictive performance comparable to, and numerically slightly better than, hypothesis-driven modeling based on a one-compartment PK model. The hypothesis-based PK/PD analysis was conducted using three model structures: direct effect, effect compartment, and indirect response. The results of these analyses, including the estimated clinical dose, were compared with those obtained from the deep-learning approach. The deep-learning model demonstrated predictive performance similar to that of the effect compartment and indirect response models and yielded a comparable estimate of the effective dose across the two modeling approaches. A simulation-based sensitivity analysis was conducted to evaluate the robustness of the dose selection derived from the deep-learning-based modeling approach. The outcomes of the analysis suggest that effective implementation of the MIDD paradigm may be enhanced through the complementary use of deep-learning-based approaches alongside established hypothesis-driven, mechanistic models, thereby supporting evidence-based drug development and regulatory decision-making.