Machine learning to predict microbial community traits driving carbon fixation

Microbial communities are ubiquitous influencers of macroscopic environments, yet overwhelming complexity makes it difficult to decipher functional relationships between specific microbes and ecosystem properties. Integrating advances in DNA sequencing technology with computational approaches like machine learning (ML) could address this problem. In [Thompson et al, PLoS One, 2019], we applied neural networks, random forest models, and indicator species analyses to correlate microbiome data (16S rRNA gene profiles) with dissolved organic carbon (DOC) content after 44 days of plant litter decomposition. We analyzed 300+ soil microcosms, including 1709 total bacterial operational taxonomic units (OTUs), and performed multiple-model feature reduction. We are now leveraging Bayesian network structure learning to infer mechanistic interactions between microbial species abundance and DOC. After training a Bayesian network model using pine litter data, we predict DOC results of independent oak litter experiments and demonstrate that relationships between microbial species abundance and DOC are conserved across litter types.