Bayesian-Guided Generation of Synthetic Microbiomes with Minimized Pathogenicity.

Area of Science:

• Microbiology
• Computational Biology
• Bioinformatics

Background:

• Multidrug resistance (MDR) poses a significant challenge in healthcare.
• Synthetic microbiomes offer a promising avenue for therapeutic interventions and research.
• Efficiently searching the vast space of synthetic microbiome variants is crucial for identifying effective candidates.

Purpose of the Study:

• To present a Bayesian optimization approach for efficient searching of synthetic microbiome variants.
• To identify synthetic microbiome candidates predictive of reduced multidrug resistance (MDR).
• To demonstrate the feasibility of creating synthetic microbiomes with customized MDR profiles.

Main Methods:

• Microbiome datasets were encoded into a low-dimensional latent space using autoencoders.
• Bayesian optimization was implemented to select synthetic microbiome variants for biological screening.
• Four acquisition functions (expected improvement, upper confidence bound, Thompson sampling, probability of improvement) were evaluated.
• Synthetic samples were prioritized based on their MDR detection.

Main Results:

• The Bayesian optimization approach enabled efficient searching over synthetic microbiome variants.
• Expected improvement, upper confidence bound, and probability of improvement acquisition functions consistently outperformed Thompson sampling.
• These methods yielded synthetic microbiome candidates with significantly fewer searches.
• The study demonstrated the creation of bespoke synthetic microbiomes with customized MDR profiles.

Conclusions:

• Combining deep latent space mapping with Bayesian learning facilitates efficient guided screening of synthetic microbiomes.
• This approach is effective for developing synthetic microbiomes tailored to specific therapeutic goals, such as reducing MDR.
• The study validates the creation of customized synthetic microbiomes for targeted applications.