Medusa: A Software Package for Construction and Analysis of Ensembles of Metabolic Networks

Genome-scale metabolic network reconstructions (GENREs) are powerful tools for predicting the metabolic behavior of organisms across broad environmental contexts and in the presence of drugs or genetic manipulation. Many applications of GENREs, such as contextualization of ‘omics data and network expansion through integration of phenotypic data, lead to multiple feasible network structures which are all equally likely. Common practice is to choose a single network and perform downstream analyses without consideration of other equally-likely networks. Recently, maintaining these alternative network structures and performing analysis in an ensemble fashion has demonstrated improved predictive performance.

Here, we present Medusa, a software package for the construction and analysis of ensembles of GENREs. Medusa is an open-source python software package that performs constraint-based reconstruction and analysis (COBRA) methods at ensemble-scale. The underlying data structures implemented in Medusa extend cobrapy, a python package for analysis of GENREs using COBRA methods.

In this presentation, we highlight three key capabilities of Medusa. The first is generation and analysis of ensembles of GENREs from phenotypic data. The second is generation of ensembles composed of alternative GENRE states created using integration of transcriptomic data. The third key capability is identification of network components associated with variable predictions throughout an ensemble using supervised and unsupervised machine learning. Using this workflow of ensemble generation and uncertainty quantification, Medusa can be used to select experiments that maximally resolve uncertainty in GENRE structure.

Simulations performed using Medusa are probabilistic in that each ensemble generates a distribution of predictions. We hope that wide adoption of ensemble analysis methods will reduce investment in GENRE-based predictions that are spurious (e.g. only one ensemble member makes the prediction) and help guide experiments to quickly improve the quality and predictive power of GENREs.