ML4SD: An Active Learning Framework for Boosting DBTL Cycles in Strain Design

The Learn phase is often the weakest link in Design-Build-Test-Learn cycles for biomanufacturing, because predictive models struggle with the combinatorial space of genetic designs such as growth-coupling gene knockouts. The talk presents ML4SD, a framework that uses gcSwarms, a binary particle swarm optimisation algorithm, to generate large and diverse design libraries, combined with active learning that balances exploration and exploitation and Shapley-based explainable AI to refine knockout recommendations and identify key metabolic interventions. In a partial validation, ML4SD-guided strains of Pseudomonas putida produced a Nylon-6 precursor at five times the wild-type yield with greater data and resource efficiency.