Microbial community differentiation and machine learning-driven optimization of nitrogen removal in multi-configured anammox systems.

Integrated fixed-film activated sludge (IFAS) reactors are widely adopted for anammox wastewater treatment, but few studies have systematically compared microbial community structure, interactions and functions among suspended sludge, pure biofilm and the two phases of IFAS. A total of 472 16S rRNA gene sequencing samples collected from 2017 to 2023 were analyzed through bioinformatic tools, diversity assessment, co-occurrence network construction and machine learning modeling. The biofilm of IFAS exhibited significantly higher alpha diversity as well as a more complex and stable microbial network with a modularity value of 0.712. Among five algorithms, LightGBM achieved the highest mean accuracy of 0.931 ± 0.034, weighted F1 score of 0.931 ± 0.035, and weighted AUC of 0.989 ± 0.009 across ten repeated stratified splits and was therefore selected for downstream SHAP analysis. SHAP analysis identified Candidatus Brocadia and Candidatus Kuenenia as primary distinguishing microbial genera. Laboratory validation identified Candidatus Anammoxoglobus, A4b, and Candidatus Brocadia as top SHAP features, with Candidatus Kuenenia enriched in biofilm (P = 0.0281), reproducing the phase differentiation observed in public datasets. Complementary shotgun metagenomic sequencing further characterized genetic potential associated with differentiation between the two phases. Fifty days of continuous operation monitoring showed stable reactor performance, suggesting possible involvement of niche complementarity in nitrogen removal. Under the tested conditions, the IFAS process showed distinct suspended-sludge and biofilm community profiles. This study provides microbial evidence and machine learning screening methods to distinguish suspended and biofilm sludge phases in lab IFAS reactors, and characterizes genetic potential associated with niche separation under the tested laboratory conditions.