Microbial biofilms in gravity-driven membrane (GDM) filtration systems can efficiently degrade the cyanotoxin microcystin (MC), but it is unclear if this function depends on the presence of MC-producing cyanobacteria in the source water habitat. We assessed the removal of MC from added Microcystis aeruginosa biomass in GDMs fed with water from a lake with regular blooms of toxic cyanobacteria (ExpL) or from a stream without such background (ExpS). While initial MC removal was exclusively due to abiotic processes, significantly higher biological MC removal was observed in ExpL. By contrast, there was no difference in MC degradation capacity between lake and stream bacteria in separately conducted liquid enrichments on pure MC. Co-occurrence network analysis revealed a pronounced modularity of the biofilm communities, with a clear hierarchic distinction according to feed water origin and treatment type. Genotypes in the network modules associated with ExpS had significantly more links to each other, indicating that these biofilms had assembled from a more coherent source community. In turn, signals for stochastic community assembly were stronger in ExpL biofilms. We propose that the less "tightly knit" ExpL biofilm assemblages allowed for the better establishment of facultatively MC degrading bacteria, and thus for higher overall functional efficiency.