CAM-Net: a context-aware network for identifying reliable microbial relations via optimal consortium.

Microbes exist within complex community contexts, particularly for key functional species whose stable colonization critically depends on specific ecological partners. However, conventional microbial correlation analyses predominantly rely on isolated pairwise metrics (e.g. Spearman, SparCC, and FlashWeave), which ignore community-level dependencies. This limitation leads to spurious associations in large-scale datasets and obscures the true structure of microbial interactions. Here, we introduce CAM-Net, a context-aware framework that identifies a target microbe's optimal consortium, a fully connected network subset that accurately predicts its abundance. By constructing networks via multi-hop information propagation, CAM-Net effectively filters false positives from indirect associations and captures complex, context-dependent patterns that are inaccessible to traditional pairwise approaches. We evaluated CAM-Net on over 25 000 human gut microbiome samples using Akkermansia muciniphila and Lactobacillus acidophilus as representatives of indigenous and transient colonizers. CAM-Net identified a coherent and reproducible consortium for A. muciniphila, but only weak association structures for L. acidophilus, consistent with their ecological behaviors. In contrast, pairwise methods produced spurious associations for both species. Notably, despite substantial geographic heterogeneity, Alistipes shahii consistently emerged as a conserved core member of the A. muciniphila consortium, demonstrating the advantage of context-aware modeling. The source code is available at https://github.com/qdu-bioinfo/CAM-Net.