From Raw Data to Optimized Models: Reinforcement Learning–Driven AI Agents to Automate Biopharmaceutical Modeling Workflows

Reinforcement Learning (RL) has recently emerged as a promising approach to automate complex decisions in machine-learning pipelines, particularly for data preprocessing and model selection. In biopharmaceutical modeling workflows such as spectroscopic analysis and batch process modeling, these decisions are typically performed manually, relying on expert judgment and extensive trial-and-error. Prior work has demonstrated that RL can effectively learn sequential data-cleaning or modeling actions, but applications to end-to-end biopharmaceutical data science workflows remain limited. In this article, an RL-driven autonomous agent is developed to optimize spectroscopic preprocessing decisions using a Markov Decision Process (MDP) formulation integrated with LangGraph. The agent learns an optimal sequence of spectral transformations including smoothing, derivative operations and model optimization that minimizes cross-validated prediction error. Using a public Near-Infrared (NIR) dataset, the proposed agent consistently discovers preprocessing chains that outperform baseline and manually developed models. This work demonstrates the feasibility of embedding RL-based decision-making within structured graph frameworks to enable reproducible, data-driven and fully automated modeling workflows.