Abstract
Anammox is a promising approach for biological nitrogen removal, but the differences in microbial community structure across different systems and their response mechanisms to environmental factors remain unclear. In this study, 206 microbial samples and 2126 environmental factor data points from three different anammox systems, including the upflow anaerobic sludge blanket (UASB), integrated fixed-film activated sludge-partial nitritation/anammox (IFAS-PN/A), and integrated fixed-film activated sludge-simultaneous nitrification, anammox and denitrification (IFAS-SNAD), were analyzed using 16S rRNA sequencing analysis, bioinformatics, and machine learning (ML) techniques. The results revealed significant differences in microbial composition among three systems, evidenced by the enrichment of Candidatus_Brocadia in IFAS-PN/A, the high-diversity community in IFAS-SNAD, and the low-diversity communities dominated by Candidatus_Kuenenia in the UASB. Co-occurrence network analysis demonstrated more tightly connected and complex interactions in IFAS-SNAD networks. Machine learning predictions further showed that the stacked model (ST-RF) achieved the highest accuracy in predicting the nitrogen removal rate (NRR), with determination coefficients (R(2)) exceeding 0.987 across all testing datasets. Moreover, SHapley Additive exPlanations (SHAP) analysis based on the stacked model revealed that the influence of key environmental factors on NRR varied by system type. These results suggested that NRR of different systems depended on the control of key environmental factors, while the significance of these environmental factors was determined by the type of system. Overall, this study enhanced the ecological and functional understanding of anammox-based processes and provided a data-driven framework for optimizing mainstream nitrogen removal.