💻 computer science

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

This paper presents an extended Human-in-the-Loop Bayesian Optimization framework that integrates constrained and robust optimization into a multi-objective Pareto front, enabling domain experts to interactively visualize and select optimal bioprocess conditions by balancing predicted performance, uncertainty, constraint satisfaction, and robustness against input perturbations.

Original authors: Samuel Stricker, Claus Wirnsperger, Alessandro Butté, Laura Helleckes, Gonzalo Guillén Gosálbez, Antonio del Rio Chanona, Mehmet Mercangöz

Published 2026-06-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Samuel Stricker, Claus Wirnsperger, Alessandro Butté, Laura Helleckes, Gonzalo Guillén Gosálbez, Antonio del Rio Chanona, Mehmet Mercangöz

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find the perfect recipe for a very expensive, delicate cake. But there's a catch: you can only bake a few cakes before you run out of money, and you don't have a perfect recipe book. You have to guess, bake, taste, and then guess again based on what you learned.

This paper describes a new way to help experts make those guesses, specifically for making medicines using living cells (bioprocesses). The authors call their method "Human-in-the-Loop Bayesian Optimization."

Here is how it works, broken down into simple concepts:

1. The Problem: The "Black Box" Cake

In the old days, scientists used a rigid checklist (like a grid) to test every possible combination of ingredients. This is slow and wasteful.
Bayesian Optimization (BO) is a smarter way. It's like having a "magic guesser" (a computer model) that learns from every cake you bake. After each bake, the model updates its map of the world, telling you where the next best guess might be.

However, traditional BO usually just says: "Bake this one specific cake." This is risky because:

The model might be wrong.
The cake might fail a safety check (like having too much sugar).
The cake might taste great in the lab but fall apart if you change the oven temperature slightly (it's not "robust").

2. The Solution: The "Pareto Dashboard"

The authors improved the system by stopping the computer from making the final decision alone. Instead, they built a dashboard that shows the human expert a list of options, not just one.

Think of this dashboard as a menu of trade-offs. Instead of saying "Bake Cake A," the computer says:

Option 1: High chance of being delicious, but we are very unsure about the ingredients. (High Reward, High Risk)
Option 2: We are very sure about the ingredients, but it might be just "okay." (Low Risk, Low Reward)
Option 3: It's delicious, but if the oven gets too hot, it will burn. (Fragile)
Option 4: It's delicious, and it won't burn even if the oven fluctuates. (Robust)

The human expert looks at this menu and picks the option that fits their current goal. Maybe today they want to take a risk to find a breakthrough; tomorrow they want to be safe.

3. The Three New Superpowers

The paper adds three specific "filters" to this menu to make it useful for real-world medicine making:

The "Safety Check" Filter (Constrained Optimization):
Imagine you are making a medicine that must be 99% pure. The computer doesn't just guess the taste; it calculates the probability that the cake will pass the purity test. If a recipe has a 90% chance of passing, it stays on the menu. If it has a 10% chance, it gets crossed out, even if it tastes amazing. This ensures the human never wastes time on recipes that are likely to fail quality checks.
The "Stress Test" Filter (Robust Optimization):
In the real world, machines aren't perfect. You might set the temperature to 37°C, but the machine might actually hit 37.2°C. The computer now simulates these tiny mistakes. It asks: "If the temperature drifts a little, does this recipe still work?"
- Fragile Recipe: Works perfectly at 37.0°C but fails at 37.2°C.
- Robust Recipe: Works great at 37.0°C and still works well at 37.2°C.
  The dashboard highlights the robust ones, so the expert can pick a recipe that won't fail due to tiny errors.
The "Uncertainty Map" (Exploration vs. Exploitation):
The dashboard shows how much the computer doesn't know. If the computer is guessing wildly in a certain area, it flags those recipes as "High Information Gain." This helps the expert decide: "Do I want to bake a safe, known cake, or do I want to bake a risky one to learn something new?"

4. How It Works in Practice

The authors tested this on a computer simulation of a Chinese Hamster Ovary (CHO) cell culture. This is a complex biological process used to make antibodies (a type of medicine).

They started with a few initial experiments.
The computer generated a "Pareto Front" (the menu of trade-offs).
The human expert looked at the menu, set some rules (e.g., "I only want recipes with a 90% chance of passing purity" and "I want them to be robust against temperature changes"), and picked a batch of new experiments.
They ran the experiments, fed the results back to the computer, and the cycle repeated.

The Result

The paper claims this method helps experts find high-performing, safe, and stable recipes much faster than old methods. It turns the optimization process from a "black box" that spits out a single answer into a collaborative conversation between the computer's math and the human's experience.

In short: The computer does the heavy lifting of calculating probabilities and simulating "what-ifs," but the human gets to hold the steering wheel, choosing the path that balances risk, reward, and safety.

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

Problem Statement
Bioprocess development, particularly in biopharmaceutical manufacturing, faces significant challenges due to high-dimensional operating spaces, expensive and time-consuming experiments, intrinsic biological variability, and stringent regulatory constraints (Quality by Design, QbD). Traditional Design of Experiments (DoE) methods are often static, defined a priori, and struggle to adapt as information accumulates, making them inefficient for complex, noisy, or poorly understood systems. While Bayesian Optimization (BO) offers a data-efficient, sequential alternative by balancing exploration and exploitation, fully automated BO frameworks often lack the interpretability and flexibility required for high-risk, regulated environments. Experts frequently need to incorporate tacit knowledge regarding manufacturability, safety, and feasibility that is difficult to encode into rigid acquisition functions. Furthermore, standard BO often fails to explicitly address two critical engineering needs: satisfying hard constraints (e.g., Critical Quality Attributes or CQAs) and ensuring robustness against input perturbations (e.g., Critical Process Parameter or CPP variability).

Methodology
The paper proposes an extension to Pareto Front Guided Sampling (PFGS), a Human-in-the-Loop (HitL) Bayesian Optimization framework. Instead of returning a single automated recommendation, PFGS presents a set of non-dominated candidate experiments on a Pareto front, allowing domain experts to interactively select the next batch based on visualized trade-offs.

The core methodology involves the following components:

Surrogate Modeling: Gaussian Process (GP) models are trained on accumulated experimental data to provide posterior mean predictions ( $\mu$ ) and uncertainty estimates ( $\sigma$ ) for process outputs.
Multi-Objective Reformulation: The framework reformulates standard BO quantities and constraints into a multi-objective optimization problem. The objectives are projected onto a Pareto front and visualized via pairwise 2D projections on an interactive dashboard.
Constrained Optimization Extension: To handle CQA constraints (e.g., purity limits), the framework incorporates the posterior probability of satisfying the constraint as an explicit Pareto objective. This probability is computed analytically from the GP posterior distribution (e.g., $P(\theta_l \leq f_c(x) \leq \theta_u)$ ).
Robust Optimization Extension: To address performance degradation under input variability, the framework integrates a Monte Carlo (MC) sampling strategy. This estimates the expected lower-confidence performance over a user-defined distribution of input perturbations (CPP deviations). This robustness score is added as a fourth Pareto objective.
HitL Workflow:
- An initial GP is trained on a dataset (e.g., Latin Hypercube Sampling).
- A large set of candidate recipes is evaluated by the surrogate models.
- A multi-objective optimizer (NSGA-II) generates the Pareto front based on objectives: predicted performance, predictive uncertainty, constraint satisfaction probability, and robustness.
- The expert defines thresholds (e.g., minimum titer, minimum constraint probability) to filter the Pareto front.
- The expert manually selects a batch of experiments from the filtered candidates or delegates selection to an automated routine based on distance to an "utopia point."
- New data is collected, the GP is retrained, and the cycle repeats.

Key Contributions
The paper makes four primary contributions to the field of BO and bioprocess engineering:

Constrained PFGS: Extends the PFGS algorithm to incorporate probabilistic CQA satisfaction as an additional Pareto objective, enabling the explicit visualization of the trade-off between performance and feasibility.
Robust PFGS: Integrates robustness criteria via MC sampling, allowing candidates to be evaluated based on their resilience to expected CPP variability, rather than just nominal performance.
Interactive Visualization: Develops an interactive dashboard and workflow that renders complex, multi-dimensional trade-offs (performance, uncertainty, constraints, robustness) interpretable for domain experts, facilitating informed decision-making without requiring new acquisition function design.
Framework Reorganization: Demonstrates that existing BO objectives can be reorganized into a HitL multi-objective framework to improve interpretability and actionability, rather than relying on novel, complex acquisition functions.

Results
The framework was validated using an eight-dimensional fed-batch Chinese Hamster Ovary (CHO) cell culture simulator with three outputs: titer, aggregation, and purity.

Initial Iteration: The system successfully identified a Pareto front where experts could filter candidates based on a minimum purity probability (e.g., >60%) and minimum titer (>1000 mg/L).
Iterative Progress: Over six iterations, the Pareto front evolved to show higher predicted titers and improved constraint satisfaction probabilities. The robustness projection demonstrated that a larger fraction of candidates retained performance under CPP perturbations as the model improved.
Expert Adaptability: The case study illustrated how expert requirements could evolve; early stages prioritized information gain (high uncertainty), while later stages prioritized conservative refinement (high constraint probability, low uncertainty).
Stopping Criteria: The framework provided a principled stopping criterion: when no candidates remained that satisfied the expert's evolving requirements, indicating the region of interest was sufficiently characterized.

Significance and Claims
The authors claim that this framework addresses the "trust gap" in automated optimization by making model recommendations interpretable and actionable for domain experts. By exposing trade-offs between predicted performance, model uncertainty, probabilistic constraint satisfaction, and input robustness simultaneously, the framework supports informed allocation of experimental resources and provides a principled stopping criterion.

The paper emphasizes that the methodology is not limited to bioprocesses or Gaussian Processes; it is broadly transferable to any engineering discipline involving expensive experimentation, constrained optimization, and uncertainty-aware decision-making. The authors modestly note that the framework does not introduce a new acquisition function but rather reorganizes existing strategies to better integrate human expertise. They acknowledge limitations, including the potential for human bias in setting requirements and the dependency on the convergence of the underlying multi-objective optimizer (NSGA-II). Future work is suggested to include multi-fidelity optimization and comparisons with fully automated strategies.