Assessing key decisions for transcriptomic data integration in biochemical networks

Abstract

To gain insights into complex biological processes, genome-scale data (e.g., RNA-Seq) are often overlaid on biochemical networks. However, many networks do not have a one-to-one relationship between genes and network edges, due to the existence of isozymes and protein complexes. Therefore, decisions must be made on how to overlay data onto networks. For example, for metabolic networks, these decisions include (1) how to integrate gene expression levels using gene-protein-reaction rules, (2) the approach used for selection of thresholds on expression data to consider the associated gene as “active”, and (3) the order in which these steps are imposed. However, the influence of these decisions has not been systematically tested. We compared 20 decision combinations using a transcriptomic dataset across 32 tissues and showed that definition of which reaction may be considered as active (i.e., reactions of the genome-scale metabolic network with a non-zero expression level after overlaying the data) is mainly influenced by thresholding approach used. To determine the most appropriate decisions, we evaluated how these decisions impact the acquisition of tissue-specific active reaction lists that recapitulate organ-system tissue groups. These results will provide guidelines to improve data analyses with biochemical networks and facilitate the construction of context-specific metabolic models.

What does your cell really do? Model-based assessment of mammalian cells metabolic functionalities using omics data

“…Doing so, they also found metabolic housekeeping functions shared across all tissues and showed similarities between metabolic activities across tissues in the same organ systems. Unfortunately, the construction and analysis of such computational models is a complex and difficult task requiring expert knowledge of the tissues and modeling framework 19,20,48 . To overcome this problem, our framework successfully combines the capacity to provide mechanistic insights of network based approaches and the simplicity of enrichment analyses.…”

“…The gene score for each reaction is selected by taking the minimum expression value amongst all the genes associated to an enzyme complex (AND rule) and the maximum expression value amongst all the genes associated to an isozyme (OR rule) 53 . Note that we have recently benchmarked the influence of preprocessing methods on the definition of the set of active genes and observed that this parameter combination presented the best performance 20 .…”

StanDep: capturing transcriptomic variability improves context-specific metabolic models

“…For preprocessing one applies a threshold to the transcriptomics dataset at a gene-or enzyme-level to decide if the gene or enzyme is expressed to sufficient levels to be considered active. Previous work identified thresholding as the most influential parameter impacting model content in context-specific models (Opdam et al, 2017;Richelle, Joshi, et al, 2018). However, it remains unclear how such thresholds be decided.…”

“…cancer cell lines, tissues, etc.). For existing thresholding methods (Richelle, Joshi, et al, 2018) to accurately capture cellular states, they must be able to select housekeeping reactions, i.e. reactions associated with housekeeping genes.…”

“…However, many algorithms require that user must come up with a list of active genes or reactions. We refer to this step as preprocessing, and it involves various decisions and; among which thresholding has the largest influence on the content of the context-specific models (Opdam et al, 2017;Richelle, Joshi, et al, 2018). Different thresholding methods embed different assumptions about transcriptomics data into contextspecific models.…”