Predictive analytics of environmental adaptability in multi-omic network models

www.nature.com/scientificreports OPEN Predictive analytics of environmental adaptability in multi-omic network models received: 20 March 2015 accepted: 14 September 2015 Published: 20 October 2015 Claudio Angione & Pietro Lió Bacterial phenotypic traits and lifestyles in response to diverse environmental conditions depend on changes in the internal molecular environment. However, predicting bacterial adaptability is still difficult outside of laboratory controlled conditions. Many molecular levels can contribute to the adaptation to a changing environment: pathway structure, codon usage, metabolism. To measure adaptability to changing environmental conditions and over time, we develop a multiomic model of Escherichia coli that accounts for metabolism, gene expression and codon usage at both transcription and translation levels. After the integration of multiple omics into the model, we propose a multiobjective optimization algorithm to find the allowable and optimal metabolic phenotypes through concurrent maximization or minimization of multiple metabolic markers. In the condition space, we propose Pareto hypervolume and spectral analysis as estimators of short term multi-omic (transcriptomic and metabolic) evolution, thus enabling comparative analysis of metabolic conditions. We therefore compare, evaluate and cluster different experimental conditions, models and bacterial strains according to their metabolic response in a multidimensional objective space, rather than in the original space of microarray data. We finally validate our methods on a phenomics dataset of growth conditions. Our framework, named METRADE, is freely available as a MATLAB toolbox. As biologists would agree, there is no biology except in the light of evolution1. However, much of the uncertainty about the behavior of a microorganism is due to the lack of statistical bioinformatics methodologies for accurate measurement of adaptability to different environmental conditions and over time2,3. Approaches involving both mathematics and bioinformatics would benefit from the study of the molecular response to the adaptation. In turn, this would enable to discover the relation between the environmental (“external”) conditions and the changes in the metabolic-phenotypic networks (the “internal” environment). At the same time, it would elucidate the genotype-phenotype relationship, which is still an open problem in biology. Many molecular levels can contribute to adaptability: (i) metabolism, i.e. the set of chemical reactions taking place in a living organism; (ii) pathway structure, namely groups of biologically-related reactions with a common goal; (iii) transcriptomics and codon usage, and in general the ability to regulate the speed of transcription and translation of genes into proteins. For instance, a highly adaptive bacterium ensures that the structure of its metabolism and the pathway productivity rapidly evolve over time due to varying environmental conditions or selective pressure4. Analogously, several recent examples show the coupling of codon usage to adaptive phenotypic variation, suggesting that the genotype functionality and behavior can be derived from the analysis of the evolution in the codon usage5. Typically, the correlation between gene expression and codon bias is large for environments similar to those in which the organism evolved, and small for dissimilar environments6. Measurements of gene expression level are able to generate transcriptional profiles of microorganisms across a diverse set of environmental conditions. Databases of environmental conditions have been Computer Laboratory - University of Cambridge, UK. Correspondence and requests for materials should be addressed to C.A. (email: ) Scientific Reports | 5:15147 | DOI: 10.1038/srep15147 1 www.nature.com/scientificreports/ recently produced for several organisms, including Escherichia coli7, Clostridium8, Salmonella9, and fission yeast10. Although such resources, coupled with statistical analysis, remain key to the interpretation of measured data, they do not provide a comprehensive understanding of the resulting cellular behavior. Examples are the cases in which similar gene expression profiles may cause different phenotypic outcomes, while different environmental conditions may give rise to similar behaviors. Additionally, the actual response to a given condition is highly dependent on the multiple cellular objectives that the microorganism is required to meet11,12. Here, we explore the adaptability of E. coli by investigating experimental conditions mapped to a multidimensional objective space. To obtain a phase-space of conditions, we add the gene expression and the codon usage layers to a flux-balance analysis (FBA) framework, therefore proposing a new multi-omic model. As a first result, we are able to optimize these layers for the overproduction of metabolites of interest, predicting the short term bacterial evolution towards the optimum. Then, we present a new method to map compendia of gene expression profiles to any metabolic objective space. Since each profile is associated with a growth condition, the objective space becomes the condition phase-space, which we investigate through principal component analysis, pseudospectra, and a spectral method for community detection. To optimize these multi-omic layers, we propose a genetic multiobjective optimization algorithm that seeks the gene expression profiles such that multiple cellular functions are optimized concurrently. We use the Pareto front as a tool to seek trade-offs between two or more tasks performed by E. coli, and specifically to score the performance when the tasks are contending with one another. We simultaneously optimize tasks by finding the best gene expression profile and codon usage array. Most notably, this may permit to determine the best environmental condition in which a bacterium has to be grown in order to reach specific optimal output values from a range of objective functions chosen by the researcher. As a particular case, it is also possible to investigate the best single or multiple gene knockouts for the given set of objectives13. The paper is organized as follows. First, we define the new multi-omic model by adding layers to FBA, in order to build the level of information required for a meaningful understanding of the landscape of experimental conditions. Using this augmented FBA framework, we optimize the model and we perform a temporal analysis of bacterial evolution towards an optimal configuration using the hypervolume indicator. Then, we introduce principal component analysis, pseudospectra and community detection methods to identify conditions mapped to close regions in the phase-space. We finally derive clusters of isoadaptability computed not in an absolute fashion, but taking into account the cellular multi-omic. Our approach is validated against a compendium of growth conditions including measurements of growth r (...truncated)