Graph-based deep learning for predictions on changes in microbiomes and biogas production in anaerobic digestion systems.
Journal:
Water research
PMID:
39826399
Abstract
Anaerobic digestion (AD), which relies on a complex microbial consortium for efficient biogas generation, is a promising avenue for renewable energy production and organic waste treatment. However, understanding and optimising AD processes are challenging because of the intricate interactions within microbial communities and the impact of volatile fatty acids (VFAs) on biogas production. To address these challenges, this study proposes the application of graph convolutional networks (GCNs) to comprehensively model AD processes. GCN models were developed to predict microbial dynamics and biogas production by integrating network analyses of high-throughput sequencing data and VFA inhibition effects. The models were trained based on the responses of anaerobic digesters to organic loading rate shock, starvation, and bioaugmentation for 281 d under various feeding conditions. Shifts in microbial community composition during AD stages and feeding conditions were successfully identified using next-generation sequencing tools. Graph topological features indicated a significant coupling between VFAs and microbial families, and the hydrogenotrophic archaeal families were most frequently connected to other families or residual acids. The GCN accurately predicted microbial abundances and gas production rates, achieving a mean squared error of 0.11 and 0.01 and a coefficient of determination of 0.72 and 0.87 for the testing dataset. These results provide valuable insights into the effects of starvation and bioaugmentation on the microbiome by utilising GCNs to model anaerobic treatment processes, predict microbial dynamics, and assess reactor productivity. Our study suggests a new modelling framework for understanding and improving AD systems by considering microbial interaction networks in relation to chemical parameter information at relevant operating scales.