Artificial Metabolic Networks: enabling neural computation with metabolic networks

Metabolic networks have largely been exploited as mechanistic tools to predict the behavior of microorganisms with a defined genotype in different environments. However, flux predictions by constraint-based modeling approaches are limited in quality unless labor-intensive experiments including the measurement of media intake fluxes, are performed. Using machine learning instead of an optimization of biomass flux - on which most existing constraint-based methods are based - provides ways to improve flux and growth rate predictions. In this paper, we show how Recurrent Neural Networks can surrogate constraint-based modeling and make metabolic networks suitable for backpropagation and consequently be used as an architecture for machine learning. We refer to our hybrid - mechanistic and neural network - models as Artificial Metabolic Networks (AMN). We showcase AMN and illustrate its performance with an experimental dataset of Escherichia coli growth rates in 73 different media compositions. We reach a regression coefficient of R2=0.78 on cross-validation sets. We expect AMNs to provide easier discovery of metabolic insights and prompt new biotechnological applications.

Identifying the essential nutritional requirements of the probiotic bacteria Bifidobacterium animalis and Bifidobacterium longum through genome-scale modeling

Although bifidobacteria are widely used as probiotics, their metabolism and physiology remain to be explored in depth. In this work, strain-specific genome-scale metabolic models were developed for two industrially and clinically relevant bifidobacteria, Bifidobacterium animalis subsp. lactis BB-12® and B. longum subsp. longum BB-46, and subjected to iterative cycles of manual curation and experimental validation. A constraint-based modeling framework was used to probe the metabolic landscape of the strains and identify their essential nutritional requirements. Both strains showed an absolute requirement for pantethine as a precursor for coenzyme A biosynthesis. Menaquinone-4 was found to be essential only for BB-46 growth, whereas nicotinic acid was only required by BB-12®. The model-generated insights were used to formulate a chemically defined medium that supports the growth of both strains to the same extent as a complex culture medium. Carbohydrate utilization profiles predicted by the models were experimentally validated. Furthermore, model predictions were quantitatively validated in the newly formulated medium in lab-scale batch fermentations. The models and the formulated medium represent valuable tools to further explore the metabolism and physiology of the two species, investigate the mechanisms underlying their health-promoting effects and guide the optimization of their industrial production processes.

ΔFBA—Predicting metabolic flux alterations using genome-scale metabolic models and differential transcriptomic data

Genome-scale metabolic models (GEMs) provide a powerful framework for simulating the entire set of biochemical reactions in a cell using a constraint-based modeling strategy called flux balance analysis (FBA). FBA relies on an assumed metabolic objective for generating metabolic fluxes using GEMs. But, the most appropriate metabolic objective is not always obvious for a given condition and is likely context-specific, which often complicate the estimation of metabolic flux alterations between conditions. Here, we propose a new method, called ΔFBA (deltaFBA), that integrates differential gene expression data to evaluate directly metabolic flux differences between two conditions. Notably, ΔFBA does not require specifying the cellular objective. Rather, ΔFBA seeks to maximize the consistency and minimize inconsistency between the predicted flux differences and differential gene expression. We showcased the performance of ΔFBA through several case studies involving the prediction of metabolic alterations caused by genetic and environmental perturbations in Escherichia coli and caused by Type-2 diabetes in human muscle. Importantly, in comparison to existing methods, ΔFBA gives a more accurate prediction of flux differences.

INTEGRATE: Model-based multi-omics data integration to characterize multi-level metabolic regulation

Metabolism is directly and indirectly fine-tuned by a complex web of interacting regulatory mechanisms that fall into two major classes. First, metabolic regulation controls metabolic fluxes (i.e., the rate of individual metabolic reactions) through the interactions of metabolites (substrates, cofactors, allosteric modulators) with the responsible enzyme. A second regulatory layer sets the maximal theoretical level for each enzyme-controlled reaction by controlling the expression level of the catalyzing enzyme. In isolation, high-throughput data, such as metabolomics and transcriptomics data do not allow for accurate characterization of the hierarchical regulation of metabolism outlined above. Hence, they must be integrated in order to disassemble the interdependence between different regulatory layers controlling metabolism. To this aim, we proposes INTEGRATE, a computational pipeline that integrates metabolomics (intracellular and optionally extracellular) and transcriptomics data, using constraint-based stoichiometric metabolic models as a scaffold. We compute differential reaction expression from transcriptomic data and use constraint-based modeling to predict if the differential expression of metabolic enzymes directly originates differences in metabolic fluxes. In parallel, we use metabolomics to predict how differences in substrate availability translate into differences in metabolic fluxes. We discriminate fluxes regulated at the metabolic and/or gene expression level by intersecting these two output datasets. We demonstrate the pipeline using a set of immortalized normal and cancer breast cell lines. In a clinical setting, knowing the regulatory level at which a given metabolic reaction is controlled will be valuable to inform targeted, truly personalized therapies in cancer patients.

Data integration uncovers the metabolic bases of phenotypic variation in yeast

The relationship between different levels of integration is a key feature for understanding the genotype-phenotype map. Here, we describe a novel method of integrated data analysis that incorporates protein abundance data into constraint-based modeling to elucidate the biological mechanisms underlying phenotypic variation. Specifically, we studied yeast genetic diversity at three levels of phenotypic complexity in a population of yeast obtained by pairwise crosses of eleven strains belonging to two species, Saccharomyces cerevisiae and S. uvarum. The data included protein abundances, integrated traits (life-history/fermentation) and computational estimates of metabolic fluxes. Results highlighted that the negative correlation between production traits such as population carrying capacity (K) and traits associated with growth and fermentation rates (Jmax) is explained by a differential usage of energy production pathways: a high K was associated with high TCA fluxes, while a high Jmax was associated with high glycolytic fluxes. Enrichment analysis of protein sets confirmed our results. This powerful approach allowed us to identify the molecular and metabolic bases of integrated trait variation, and therefore has a broad applicability domain.

Stoichiometric Modeling of Artificial String Chemistries Reveals Constraints on Metabolic Network Structure

Uncovering the general principles that govern the structure of metabolic networks is key to understanding the emergence and evolution of living systems. Artificial chemistries can help illuminate this problem by enabling the exploration of chemical reaction universes that are constrained by general mathematical rules. Here, we focus on artificial chemistries in which strings of characters represent simplified molecules, and string concatenation and splitting represent possible chemical reactions. We developed a novel Python package, ARtificial CHemistry NEtwork Toolbox (ARCHNET), to study string chemistries using tools from the field of stoichiometric constraint-based modeling. In addition to exploring the topological characteristics of different string chemistry networks, we developed a network-pruning algorithm that can generate minimal metabolic networks capable of producing a specified set of biomass precursors from a given assortment of environmental nutrients. We found that the composition of these minimal metabolic networks was influenced more strongly by the metabolites in the biomass reaction than the identities of the environmental nutrients. This finding has important implications for the reconstruction of organismal metabolic networks and could help us better understand the rise and evolution of biochemical organization. More generally, our work provides a bridge between artificial chemistries and stoichiometric modeling, which can help address a broad range of open questions, from the spontaneous emergence of an organized metabolism to the structure of microbial communities.