Microbial community biomarkers can forecast methane production in full-scale anaerobic digesters.

Methane production from wastewater sludge via anaerobic digestion is a complex process and a disturbance in any one of the microbial stages can lead to eventual failure. Hence, it is desirable to detect disturbances as soon as possible. Although machine learning has been used to predict methane production from a variety of different substrates, there are no studies using metagenomic or -transcriptomic microbial community data as predictor variables. We used random forest analysis on a combination of physicochemical and microbial predictors to forecast methane production from three full-scale sludge digesters representing replicates of one another in a wastewater treatment plant in Singapore. Digesters were sampled for 25 weeks, and 42 physicochemical variables were measured along with shotgun metagenome and total RNA transcriptome sequencing. Models built using samples from a single digester yielded reactor-specific predictors, largely due to the limited sample size per reactor and the influence of rarer taxa. When data from the three digesters were combined, the best predictors included both substrate-related physicochemical parameters, such as chemical oxygen demand, and microbial taxa. Simulation using learning curves indicated that 150 to 200 samples instead of the 75 used would have yielded the most accurate methane prediction. The selection of many unidentified operational taxonomic units as microbial predictors suggests the existence of important yet unknown microorganisms in anaerobic digestion. The prediction model supports onsite digester surveillance by identifying digester-specific predictors through sufficient sampling, after which only those predictors need to be measured for subsequent monitoring.