Bayesian Optimization for Biotech R&D Portfolios: What Drug Discovery Can Learn from Hyperparameter Tuning

Bayesian optimization (BO) is a cornerstone of automated machine learning, where it is used to tune hyperparameters—such as learning rates and layer sizes—under limited computational budgets. The same mathematical framework, however, remains nearly unknown in biotech R&D portfolio management, where decision-makers must allocate finite capital across drug candidates at different stages of development. In both settings, the problem is identical: maximize an unknown objective function (model accuracy or portfolio value) given a constrained budget of experiments or investments.This review makes three contributions. First, we provide a non-technical introduction to BO, explaining Gaussian process surrogates and acquisition functions for biologists and portfolio managers. Second, we present parallel case studies: hyperparameter tuning in deep learning, reaction condition optimization in chemistry, and adaptive portfolio allocation in finance. Third, we formally map the biotech R&D portfolio problem onto the BO framework, defining the search space (drug assets), objective (expected net present value), and feedback (noisy clinical signals). A toy simulation demonstrates that BO outperforms random and greedy allocation strategies. We conclude with practical recommendations for implementing BO in biotech settings and discuss open challenges, including asset correlation and regulatory constraints.