(191di) A Deep Learning Framework Decodes Coordination of Microbial Metabolism Under Genetic and Environmental Perturbations | AIChE

(191di) A Deep Learning Framework Decodes Coordination of Microbial Metabolism Under Genetic and Environmental Perturbations

Authors 

Oyetunde, T. - Presenter, Washington University in St. Louis
Czajka, J., Washington University in St. Louis
Tang, Y., Washington University in St. Louis
One of the main goals of systems biology research is to decode the complex relationship between an organism’s genotype and phenotype. A key method that has been used extensively towards this end is inferring regulatory structure and active regulatory events from biological (omics) data. Metabolism ties extracellular information with intracellular signals in a complex regulatory network made up of modular regulatory circuits built on a small set of metabolites. It has been hypothesized that metabolic regulation exists as a nested loop of interacting modules with local and global levels of regulation. Key challenges identified with omics data integration include sparse and disparate datasets, computational complexity, and difficulty in deducing biological meaning from the integration.

We present a novel framework, DeepOmics, that attempts to solve these challenges by uniquely integrating data-driven and constraint-based techniques. DeepOmics is a specially designed multi-hidden-layer neural network, in which each neuron and its associated weights and biases are designed to mimic actual biological regulation. Thus, DeepOmics goes beyond mere prediction of phenotypes but also seeks to uncover regulatory interactions between genes, gene products and the associated metabolic network. This is achieved by leveraging machine learning techniques of deep learning and transfer learning in conjunction with well-established constraint-based methods.

One key application of our framework, that distinguishes it from previous attempts, is the characterization of the efficiency of metabolic engineering strategies. DeepOmics builds on established computational strain design algorithms to uncover influential factors (such as medium composition, cultivation modes, pathway steps, and genetic engineering strategies) that are critical to the principal metrics of effectiveness: yield, titer, and rate. In this way DeepOmics enables efficient screening of metabolic engineering designs based on decades of data from published literature.

We have tested DeepOmics on the central metabolism of the microbe Escherichia Coli. Our framework recapitulates known biology and generates testable predictions about cellular regulatory interactions involved from genotype to phenotype. DeepOmics can provide valuable insights that will enhance applications in biotechnology, human health, and environmental remediation

References

  1. Machado, D., Herrgard, M. J. & Rocha, I. Stoichiometric Representation of Gene-Protein-Reaction Associations Leverages Constraint-Based Analysis from Reaction to Gene-Level Phenotype Prediction. PLoS Comput. Biol. 12, 1–24 (2016).
  2. Gerosa, L. et al. Pseudo-transition Analysis Identifies the Key Regulators of Dynamic Metabolic Adaptations from Steady-State Data. Cell Syst. 1, (2015).
  3. Chubukov, V., Gerosa, L., Kochanowski, K. & Sauer, U. Coordination of microbial metabolism. Nat. Rev. Microbiol. 12, 327–40 (2014).