Large Language Model-Enhanced MFSSG Enables Effective and Accurate Integration Analysis for Microbiome−Metabolome Data

The massive volume of paired microbiome and metabolome measurementsgenerated through high-throughput sequencing and mass spectrometryhas created an urgent need for dedicated, integrative analytical pipelines. Existing tools lack the capacity to simultaneously support sample clustering, multi-omics visualization, disease risk stratification, and biological age quantification using matched dual-modal datasets. To overcome this analytical bottleneck, we introduce the matrix factorization with shared and specific graph learning (MFSSG) toolkit: an unsupervised matrix factorization framework that jointly learns both shared and modality-specific graph structures. MFSSG preserves cross-omics spatial consistency by leveraging shared nearest-neighbor graphs derived from analytical profiles and incorporates two complementary regularization termsentropy regularization to enhance cluster discriminability and a pre-trained microbial graph Laplacianto mitigate measurement bias inherent in microbiome sequencing versus metabolomic mass spectrometry data. Validated across three independent multi-omics cohorts derived from gastric and gut biospecimensincluding gastric cancer, end-stage renal disease, and type 1 diabetesMFSSG consistently outperforms state-of-the-art integration methods in disease classification. Furthermore, a novel pairwise learning module enables parallel and accurate quantification of the metabolic age and microbial age, facilitating refined patient risk stratification. This versatile toolkit provides a robust chemometric strategy for uncovering latent biological patterns embedded within heterogeneous multi-omics analytical data.