Evaluating the impact of clinical trial design features on probability-of-success prediction

Estimating the probability of success (POS) of clinical trials remains a major challenge in drug development. Legacy forecasting approaches in the biopharmaceutical industry rely on historical averages stratified by development phase and therapy area, which obscures the contribution of trial design to clinical outcomes. Here we test the hypothesis that clinical trial design characteristics provide independent predictive signal that can improve estimation of development outcomes. Using a longitudinal dataset of 35,662 clinical trial transitions spanning 27,900 investigational drugs and 5,001 trial sponsors, we developed phase-specific machine learning-based models evaluated under time-aware validation. Models incorporating trial design features produced consistent directional improvements in Phases 1 and 2 and clearer gains in Phase 3 compared with models relying only on drug, indication, and sponsor attributes. Feature importance analysis revealed that trial design-related variables ranked among the strongest predictors when available, whereas exclusion of these variables shifted model reliance toward sponsor experience and investigational drug characteristics. These findings challenge continued industry reliance on aggregate historical benchmarks. They support a shift toward feature-resolved, data-driven models in which clinical development success is predicted from the interaction of trial design, sponsor capability, and investigational drug characteristics. Further, the proposed framework establishes a scalable approach for integrating emerging clinical development data sources that may enable more informed portfolio strategy and trial design decisions.