Simulated metabolic profiles reveal biases in pathway analysis methods.

Abstract

Introduction: Initially developed for transcriptomics data, pathway analysis (PA) methods can introduce biases when applied to metabolomics data, especially if input parameters are not chosen with care. This is particularly true for exometabolomics data, where there can be many metabolic steps between the measured exported metabolites in the profile and internal disruptions in the organism. However, evaluating PA methods experimentally is practically impossible when the sample's "true" metabolic disruption is unknown.

Objectives: This study aims to show that PA can lead to non-specific enrichment, potentially resulting in false assumptions about the true cause of perturbed metabolic states.

Methods: Using in silico metabolic modelling, we can create disruptions in metabolic networks. SAMBA, a constraint-based modelling approach, simulates metabolic profiles for entire pathway knockouts, providing both a known disruption site as well as a simulated metabolic profile for PA methods. PA should be able to detect the known disrupted pathway among the significantly enriched pathways for that profile.

Results: Through network-level statistics, visualisation, and graph-based metrics, we show that even when a given pathway is completely blocked, it may not be significantly enriched when using PA methods with its corresponding simulated metabolic profile. This can be due to various reasons such as the chosen PA method, the initial pathway set definition, or the network's inherent structure.

Conclusion: This work highlights how some metabolomics data may not be suited to typical PA methods, and serves as a benchmark for analysing, improving and potentially developing new PA tools.