Interpretable Machine Learning Reveals Integrated Water Chemistry and Parameter-Specific Nonlinear Responses Shaping Legionella spp. and Mycobacterium spp. in Drinking Water

Traditionally, studies have explored the impacts of individual water chemistry parameters on the persistence of Mycobacterium spp. and Legionella spp. in isolation with the underlying assumption that these associations are likely monotonic in nature. Yet chemical and microbiological changes are complex, and associations are likely highly combinatorial. In this study, we use interpretable machine learning models to disentangle the integrative and nonlinear associations between water chemistry and occurrence/abundance of Mycobacterium spp. and Legionella spp. Seasonal data from source water, point-of-entry and distribution systems of eight full-scale drinking water systems demonstrated that shifts in overall water chemistry were associated with the changes in microbial abundance during treatment and distribution. Machine learning models indicated moderate predictive ability of integrated water chemistry towards Legionella spp. abundance and towards the occurrence of both Legionella spp. and Mycobacterium spp., whereas predictive performance for Mycobacterium spp. abundance was limited. The association between nitrate and Legionella spp. abundance was disinfectant regimes dependent, while dissolved organic carbon exhibited a concentration dependent response type (i.e., positive and negative association). In chloraminated systems, Legionella spp. abundance was positively associated with ammonia and nitrate, highlighting the critical role of nitrification. Here, it appears that pH likely influences the initial colonization of Legionella spp. while ammonia governs its abundance in drinking water. Overall, this study demonstrates that integrated water chemistry and parameter-specific nonlinear effects collectively explain persistence of Mycobacterium spp. and Legionella spp. in drinking water systems.