Metabolic engineering aims to cost-effectively produce a molecule of interest by integrating diverse genetic manipulation strategies. However, traditional manual experimentation can explore only a fraction of the extensive combinatorial space of possible genetic modifications. To overcome this limitation, we have developed a self-driving biofoundry that integrates synthetic biology, laboratory automation, and machine learning to engineer microbial cell factories. Specifically, we utilized a tri-functional multiplexed CRISPR-AID system capable of simultaneous transcriptional activation, interference, and gene deletion, effectively perturbing metabolic networks. This approach was integrated into an automated workflow at the Illinois Biological Foundry for Advanced Biomanufacturing (iBioFAB), streamlining plasmid assembly, yeast transformation, genetic manipulation, and high-throughput screening. As a proof-of-concept, we introduced the beta-carotene biosynthetic pathway into Saccharomyces cerevisiae, targeting 30 genes to investigate their combinatorial effects, thereby generating a combinatorial space exceeding millions of potential variants. To efficiently navigate this vast genetic space, we developed a machine learning algorithm utilizing graph neural networks to extract metabolic perturbation features. An active learning strategy was employed to iteratively train and predict optimal genetic modifications for subsequent experimental cycles. This closed-loop design-build-test-learn (DBTL) cycle dynamically guided experimental designs, significantly enhancing the efficiency of metabolic engineering in microbial cell factory development. We anticipate this self-driving biofoundry will represent a new paradigm in constructing microbial cell factories, significantly enhancing scalability, efficiency, and innovation in industrial biomanufacturing.
