Modern biomanufacturing relies on iterative Design-Build-Test-Learn (DBTL) cycles, yet the "Learn" phase is frequently the weakest link due to the limited predictive power for complex biological systems. Conventional approaches are unable to cope with the "combinatorial explosion" of the genetic design space, particularly when engineering growth-coupled phenotypes through gene knockouts.

In this presentation, we propose ML4SD, a computational framework that integrates machine learning into the DBTL cycle with the objective of accelerating strain design. The present approach utilizes gcSwarms, a novel binary particle swarm optimisation algorithm that generates extensive and diverse design libraries with a view to enhancing machine learning model generalization. The employment of an active learning strategy that balances exploration and exploitation enables ML4SD to iteratively refine knockout recommendations whilst identifying key metabolic interventions through the utilization of Shapley-based explainable AI.

A partial validation of ML4SD was made using Nylon-6 precursor production in Pseudomonas putida, resulting in a yield that was five times higher than that obtained with the wild type. In addition, ML4SD has been shown to exhibit superior data- and resource-efficiency in comparison to its data source generation. Overall, this methodology establishes a robust and user-friendly platform for the autonomous development of high-performing microbial cell factories.