Local Sample Cohesion Normalization: Preserving Inherent Biological Heterogeneity in Metabolomics Data

Metabolomics data from biofluids like urine or cell cultures are frequently confounded by unwanted overall sample concentration (or dilution effects). Conventional normalization methods, such as Constant Sum Normalization (CSN), L2-Norm Normalization (L2N), Probabilistic Quotient Normalization (PQN), and Quantile Normalization (QT), rely on a unified global reference, failing to account for inherent biological heterogeneity (e.g., interindividual variability, subgroup divergences). This limitation can distort biological data structures and compromise downstream analyses. To address this issue, we developed Local Sample Cohesion Normalization (LSCN), that corrects dilution effects while preserving biological heterogeneity. LSCN constructs a sample-specific neighbor set for each spectrum based on pairwise similarity in a reduced-dimensional space and performs locally weighted normalization within these neighborhoods. This approach mitigates technical bias without enforcing artificial global uniformity. We rigorously validated LSCN against CSN, L2N, PQN, and QT normalization using simulated data sets with known heterogeneity and diverse real-world metabolomics data sets (urine, cells, tissues, tea leaves). LSCN demonstrated superior performance in Preserving heterogeneity, achieving significantly higher global and local structural similarity to ground-truth references; Retaining biological signals, enhancing identification of differential metabolites, correlation networks, pathway enrichment, and classification accuracy; and Effectively correcting dilution effects, yielding more accurate normalization factors and reduced within-group variance. LSCN offers a robust, biologically faithful preprocessing framework for metabolomics, improving reliability in downstream analyses.