# Are the tools fit for purpose? Network inference algorithms evaluated on a simulated lipidomics network

## TL;DR

This paper evaluates how well different algorithms can infer metabolic networks from simulated data, highlighting the challenges in accurately recovering network structures.

## Contribution

The study benchmarks network inference algorithms using a simulated lipidomics network to assess their effectiveness under varying sample sizes.

## Key findings

- Network inference faces significant challenges even with large sample sizes and simulated data.
- Correlation-based methods can distinguish between different metabolic states but not direct pathways.
- Results suggest utility in identifying broad metabolic changes rather than specific interactions.

## Abstract

Various methods have been proposed to construct metabolic networks from metabolomic data; however, small sample sizes, multiple confounding factors, the presence of indirect interactions as well as randomness in metabolic processes are of major concern.

In this study, we benchmark existing algorithms for creating correlation- and regression-based networks of changes in metabolite abundance and evaluate their performance across different sample sizes of a generative model. Using standard interaction-level tests and network-scale analyses based on centrality scores, we assess how well these methods recover represented metabolomic networks. Our findings reveal significant challenges in network inference and result interpretation, even when sample sizes are significant and data are the result of computer modeling of metabolic pathways. Despite these limitations, we demonstrate that correlation-based network inference can, to some extent, discriminate between two different metabolic states in the computational model. This suggests potential utility in distinguishing overarching changes in metabolic processes but not direct pathways in different conditions.

All relevant data is provided at https://github.com/TheCOBRALab/metabolicRelationships