Topic modeling reveals thermally partitioned and taxonomically distinct microbial subcommunities across prokaryotes and phytoplankton in the Laurentian Great Lakes

Identifying discrete microbial assemblages and their environmental drivers across multiple biological fractions simultaneously remains a central challenge in aquatic microbial ecology. We applied an integrated analytical pipeline built around Latent Dirichlet Allocation (LDA) to an eight-year 16S rRNA amplicon time series from the Laurentian Great Lakes, spanning four size-fractionated biological blocks free-living prokaryotes, particle-associated prokaryotes, and small and large chloroplast-containing eukaryotes. LDA resolved ecologically coherent subcommunities whose taxonomic identity was consistently defined at the order and class level, with fingerprint taxa confirmed by discriminant analysis. Shannon entropy differences between blocks reflected fundamental differences in dispersal capacity and environmental filtering free-living prokaryotes and large eukaryotes showed higher mixing than particle-associated prokaryotes and small eukaryotes. Temperature dominated environmental structuring across all blocks, assessed through Limma and random forest with SHAP, with secondary drivers differing by size fraction. Group Compositional Analysis jointly integrating all four blocks revealed that thermal stratification and lake chemistry organize microbial communities coherently across all size fractions simultaneously. Warm stratified and cold inversely-stratified waters harbored largely non-overlapping assemblages across all four blocks, with cold water specialists - including chemolithotrophic deep-branching lineages and silica-dependent diatoms, having no warm water equivalents.