Machine Learning assisted engineering of artificial microbial consortia for the delivery of bio-plastic materials from renewable waste

Project team: Leonidas Matsakas (project Leader – WISE), Christakopoulos Paul, Anjali Purohit, Ulrika Rova

Partners: Marcus Liwicki (project Leader – WASP), Machine Learning, LTU

Duration: 2024

Funded by the Wallenberg AI, Autonomous Systems and Software Program (WASP) & Wallenberg Initiative Materials Science for Sustainability (WISE)

The project focuses on engineered artificial microbial consortia that provide the potential to design microorganisms that can engage in symbiotic relationships, effectively bypassing the metabolic constraints of single strains and the inefficiencies of natural consortia. Traditional techniques to build artificial consortia are labor-intensive, resource-heavy, and often neglect the intricate microbial interactions, which can lead to unexpected outcomes or even consortium collapse. The recent advances in high-throughput screening, sequencing, computational techniques, and machine learning (ML) have paved the way for designing, building, and analyzing complex microbial consortia. This momentum has fostered a new realm in synthetic biology and biotechnology. The goal is to create artificial communities that mirror the robustness, efficiency, and flexibility of natural systems. This evolution has been propelled by high-throughput biological data sources, encompassing transcriptomics, proteomics, and metabolomics, to name a few. One notable application of these techniques is the development of a machine learning-based Automated Recommendation Tool (ART) for synthetic biology. This tool offers a set of suggested

strains for subsequent engineering cycles, coupled with probabilistic predictions of their production metrics.

From a machine learning and data science perspective, ART is very well conceptualized and easy to use, but does not include the most recent accomplishments in meta learning and learning with sparse data (SSL – Self supervised learning), which would be useful for scenarios where real experimental data is costly – and many particular solutions exist which have a similar high-level structure.