The cNED framework: modeling core microbial communities and metabolic functions with SHAP-interpretable environmental...

Microbial communities in river ecosystems regulate biogeochemical cycling and serve as natural bioremediators for environmental pollutants. However, accurately predicting their dynamic responses to changing conditions remains a significant scientific challenge because of the complexity of microbial interactions and ecosystem-scale feedbacks. Here, a novel compositional neural encoder-decoder (cNED) framework was developed, coupling environmental variables with microbial profiles based on an extensive collection of 473 samples from the upper Yangtze River. A total of 157 core bacterial OTUs were identified from 27,932 OTUs by the occupancy-frequency method, which were predominantly governed by deterministic assembly processes. The identified core microbiome demonstrated significant functional associations with carbon and nitrogen cycling. Compared with conventional modeling approaches (multilayer perceptron, random forests, linear regression), the cNED framework demonstrated superior predictive performance, achieving high accuracy in taxonomic prediction and functional profile prediction (carbon cycling: R² = 0.85; nitrogen cycling: R² = 0.52). The Shapley additive explanations (SHAP) analysis identified spatial gradients and temperature as key environmental drivers. Generalized Additive Models uncovered phylum- and function-specific tipping points: Proteobacteria exhibited a dual-threshold thermal niche (20.5-27 °C), while functions like methylotrophy and nitrogen fixation responded nonlinearly to temperature and TN, revealing unimodal or monotonic transitions. The cNED framework developed in this study establishes an interpretable predictive framework for forecasting microbial community and functional responses to environmental perturbations, offering valuable insights for evidence-based river ecosystem management and climate adaptation strategies.