Bioprocess optimisation via joint machine learning and metabolic modelling

Optimisation of the design-build-test-learn cycle remains a bottleneck for developing and efficiently manufacturing the next generation of bioproducts. To address this challenge, research has traditionally focused on either data-driven or mechanistic modelling, but the emerging consensus in the field highlights that their strategic integration, often termed hybrid modelling, offers significant advantages, particularly for complex systems where data is sparse. Here, we introduce a hybrid framework where these two methodological lineages are combined to frame development parameters within the metabolism of a production system, therefore achieving mechanism-informed predictions for follow-up experiments. We present three scenarios that exemplify how this framework can be leveraged to guide and accelerate the development of novel bioprocesses, even with small datasets available. We validate our framework by applying it to two heterologous peptide production scenarios in Escherichia coli, explored through commonly relevant experimental factors such as inducer concentration and production strain. Using our framework, we identify key metabolic pathways and reactions that contribute to productivity and whose activity is modified by individual experimental factors like temperature and plasmid used. Furthermore, we show that the biological patterns extracted from this hybrid approach can complement the experimental design, informing predictive models of process performance. Our approach is general and can be tailored to a large array of processes, and thus holds potential for boosting both proof-of-concept and industrial projects, contributing to more efficient and sustainable biomanufacturing.