**Background:** Metabolic profiling (metabolomics) aims to measure small molecules in biofluids to characterize disease states, but no single analytical technique can measure the entire metabolome. Designing the right experiment is time-consuming and resource-intensive. Computational methods can help by predicting which metabolites are likely to be altered in a given condition, guiding experimental design and data interpretation. Constraint-based modelling (CBM) of genome-scale metabolic networks can simulate metabolic fluxes, and previous work used Flux Variability Analysis (FVA) to predict biomarkers as binary (changed/not changed) based on shifts in flux bounds. However, FVA does not capture the most frequent flux values or allow ranking of metabolite changes by intensity.
**Methods:** The authors developed SAMBA (SAMpling Biomarker Analysis), which uses random sampling (optGpsampler algorithm) to generate flux distributions for all exchange reactions (metabolite import/export) in a genome-scale metabolic network under two conditions: a wild-type (WT) state and a mutant (MUT) state simulating a metabolic perturbation (e.g., gene knockout). For each metabolite's exchange reaction, a z-score is calculated by comparing the WT and MUT flux distributions: z_i = mean(MUT_i - WT_i) / std(MUT_i - WT_i). Metabolites are then ranked by the absolute value of their z-score (SAMBARank), with higher ranks indicating greater predicted change in import/export flux. The method was tested on two models: Recon 2 (7,440 reactions) for inborn errors of metabolism (IEMs) and Human1 (13,024 reactions, 8,363 metabolites) for mGWAS data. For IEMs, 49 diseases were simulated by knocking out the responsible gene(s), and predictions were compared to a curated biomarker dataset. For mGWAS, SNPs associated with metabolic traits (e.g., SCD and ACADS) were simulated as gene knockouts, and predicted metabolite ranks were compared to experimentally significant metabolites from Suhre et al. (2011). Statistical significance was assessed using hypergeometric tests, and chemical class enrichment was performed using BiNChE and ChemRich.
**Key Results:** For Xanthinuria type I (XDH knockout), FVA failed to predict the expected decrease in urate (bounds unchanged), while SAMBA correctly showed a decreasing shift in the sampling distribution. For hypoxanthine, both methods predicted the expected increase. Across 49 IEMs, SAMBA achieved a maximum accuracy of 80% (at rank cutoff 110) for predicting the correct direction of change, slightly higher than the FVA-based accuracy of 76.8%. For the SCD SNP (rs603424), four out of five experimentally significant metabolites were present in Human1, and SAMBA predicted the correct direction of change for all four, with ranks in the top 13% (two in the top 1%). FVA incorrectly predicted an increase for myristoleate (expected decrease). For the ACADS SNP, carnitine (rank 22) and several amino acids were highly ranked, consistent with known metabolic roles. Hypergeometric tests for SCD showed that finding ~6 expected metabolites in the top 100 predictions was highly significant (p-value << 0.01). Chemical class enrichment of the top 10 SAMBA predictions for SCD revealed all were lipids (8 fatty acids), consistent with SCD function. Enrichment of the top 50 predictions using ChemRich identified additional classes like HETE (hydroxyeicosatetraenoic acids), cholestenes, and cholestadienols, which were not detected using only the experimentally significant metabolites.
**Clinical Implications:** SAMBA provides a ranked list of metabolites predicted to be differentially abundant in biofluids under a given metabolic perturbation. This can be used upstream of experiments to select the most suitable analytical platform (e.g., lipidomics if lipids are predicted), or downstream to guide targeted identification of metabolites from raw data (e.g., selecting standards for level 1 annotation). The method can also suggest novel metabolites or chemical classes for further study, potentially uncovering underappreciated pathways. For example, in SCD, SAMBA highlighted odd-chain fatty acids and HETE as potential targets. The approach is applicable to genetic diseases, and can be extended to simulate knock-downs (e.g., drug effects) or other perturbations. Limitations include the reliance on model accuracy, the computational cost of sampling, and the difficulty of evaluating false positives when metabolites are not measured experimentally. The authors note that the method is implemented in a freely available workflow.