Closed-loop optimization of high-throughput robotic platforms for reproducible bioprocess development

Abstract

To confidently accelerate bioprocess development using high-throughput roboticplatforms, parallel experiments should be reproducible whereas the generated data mustbe FAIR. To this aim, higher levels of automation for planning, redesigning, executing,and monitoring of related experimental-computational workflows are needed. In thiswork, an automation hierarchy for closed-loop optimization based on Bayesian update,belief propagation and Thompson sampling is proposed. For online redesign of paralleldynamic experiments, asynchronous rolling out beliefs about beliefs based on simulateddata is used for counterfactual assessment of the effect of alternative (re)design decisionson the distributions of model parameters. A computational implementation of theBayesian automation hierarchy using Apache Airflow is described. Simulation resultsobtained using an E. coli case study are presented to argue that parallel dynamicexperiments made using the proposed Bayesian hierarchy are reproducible whereasexperimental-computational workflows and data are FAIR by design.