Local neighbor Normalization: Reconciling accurate normalization and heterogeneity recovery in large-scale...

Metabolomics studies often grapple with the dilution effect, where sample concentrations vary due to inconsistent handling or biological diversity, particularly in samples like urine, saliva, or cell extracts. This variation can mask true metabolic differences, complicating data interpretation. Traditional normalization methods, such as Constant Sum Normalization (CSN), Probabilistic Quotient Normalization (PQN), and Maximal Density Fold Change (MDFC), assume that all samples share a certain invariant statistic and overlook data heterogeneity, potentially erasing the dataset's heterogeneity essential for distinguishing biological subgroups.

To address this, we introduce Local Neighbor Normalization (LNN), a novel approach that corrects for dilution effects while preserving the intrinsic variability of metabolomics data. LNN identifies a neighbor set for each sample based on similarity metrics and normalizes each sample against a tailored reference spectrum derived from these neighbors. Through comprehensive evaluations on both simulated and real metabolomics datasets from NMR, GC-MS, and LC-MS platforms, LNN demonstrated superior performance over CSN, PQN, and MDFC. Specifically, it achieved better elimination of dilution effects, recovery of inter-sample heterogeneity and inter-metabolite correlations, as evidenced by metrics such as the D-statistic and correlation recovery rates. Notably, LNN excels in datasets with over 50 % differential metabolites, safeguarding local data structures critically for downstream analyses like biomarker discovery.

LNN constructs sample-specific reference spectra based on a local neighbor set. This approach ensures that normalization accounts for dilution effects without compromising local structure of the data, which is crucial for biological interpretation. Additionally, LNN demonstrates superior performance in recovering inter-sample heterogeneity and metabolite correlations, especially in datasets with high proportions of differential metabolites. This method's versatility, robustness against noise, and applicability across various metabolomics platforms make it a significant advancement in the field.