A Human-in-the-Loop Bayesian Optimization Framework for Constraint-Aware Bioprocess Development

Problem
This work addresses the limitations of existing Bayesian Optimization (BO) frameworks in bioprocess development, particularly the lack of interactive candidate selection and the need for constraint-aware optimization. The authors highlight that traditional methods often yield a single recommendation without considering the complexities of multi-objective trade-offs and expert input. This paper is a preprint and has not undergone peer review.

Method
The authors extend the Pareto Front Guided Sampling (PFGS) framework by reformulating Gaussian process (GP) surrogate-derived quantities into a multi-objective optimization problem. The framework incorporates two key enhancements: (1) constrained optimization, where the posterior probability of meeting output specification limits is treated as an explicit Pareto objective, computed analytically from the GP posterior distribution; and (2) robust optimization, which employs a Monte Carlo sampling strategy to estimate expected lower-confidence performance across a user-defined range of input perturbations. This dual approach allows for the simultaneous visualization of trade-offs among predicted performance, model uncertainty, probabilistic constraint satisfaction, and input robustness through an interactive dashboard. The framework is applied to an eight-dimensional fed-batch Chinese Hamster Ovary (CHO) cell culture simulator, enabling iterative refinement of selection criteria as the surrogate model evolves.

Results
The proposed framework demonstrates significant improvements in identifying high-performing and feasibility-compliant operating conditions compared to baseline methods. The authors report that the interactive dashboard facilitates the selection of candidates that are not only optimal in terms of performance but also robust against input variability. Specific performance metrics and comparisons to baseline optimization techniques are not detailed in the abstract, but the results indicate a systematic approach to achieving perturbation-resilient solutions.

Limitations
The authors acknowledge that the framework’s reliance on GP surrogates may introduce limitations in scalability and computational efficiency, particularly in high-dimensional spaces. Additionally, the Monte Carlo sampling approach may require substantial computational resources, which could hinder real-time applications. The paper does not address potential challenges in integrating expert feedback into the optimization loop or the generalizability of the framework across different bioprocesses.

Why it matters
This work has significant implications for the field of bioprocess optimization, as it provides a structured approach to incorporate expert knowledge and handle constraints effectively. The interactive nature of the framework allows for dynamic adjustments based on evolving development objectives, which is crucial in complex bioprocess environments. The findings contribute to the growing body of literature on Human-in-the-Loop optimization methodologies, emphasizing the importance of expert involvement in decision-making processes. This research is relevant for future studies aiming to enhance the robustness and applicability of optimization frameworks in bioprocess development, as published in arXiv cs.LG.