Ensemble Approaches for Enhancing Robustness and Reproducibility in RNA-Seq-Based Genome-Scale Metabolic Models

Genome-scale metabolic models (GEMs) can be (re)constructed and contextualized using RNA sequencing data. However, the reconstruction process is sensitive to data variation and heavily relies on the methods used to quantify the RNA-Seq data, as well as algorithms to reconstruct the model. Consequently, this may lead to substantial differences or even contradictions between the resulting models. To address this issue, we propose an ensemble approach that accounts for uncertainty in model structure due to RNA-Seq data variability and different RNA-Seq quantification methods. The ensembles of models serve multiple purposes. First, we demonstrate how different RNA-Seq quantification methods can influence the resulting model content. We illustrate how the ensemble of models can guide the selection of appropriate thresholds for RNA-Seq data to mitigate this effect. Second, we leverage the information contained in the ensemble with machine learning to identify reactions that are robust to uncertainty in RNA-Seq quantification discrepancies and data variation. The strength of the ensemble lies in its ability to recognize reactions that are present in only a few models representing a particular condition, but are entirely absent in others. These reactions can serve as markers of metabolic differences between conditions, despite their low abundance. Third, the ensemble accounts for the uncertainty encountered in several steps of the context-specific metabolic model reconstruction process. Consequently, it can guide and standardize (manual) curation efforts, improve robustness and reproducibility of models, and highlight relevant metabolic derangements in the context of disease. We illustrate our approach on metabolic rewiring in pro-inflammatory endothelial cells.