Hybrid models enabling neural computations with metabolic networks

Abstract

AbstractConstraint-based mechanistic models have largely been exploited to predict the phenotype of microorganisms in different environments. However, phenotype predictions are limited in quality unless labor intensive experiments including the measurement of media uptake fluxes, are performed. We show how hybrid - mechanistic and neural – models provide ways to improve phenotype predictions. Our hybrid models named Artificial Metabolic Networks (AMNs) surrogate constraint-based modeling, make metabolic networks suitable for backpropagation and, consequently, can serve as an architecture for machine learning. We first show how learning principles brought by AMNs can replace the optimization principle of constraint-based modeling with excellent performances for variousin silicotraining sets. We then illustrate how AMNs outperform mechanistic models withEscherichia coligrowth rates measured in 110 different media compositions reaching regression coefficients > 0.76 on cross-validation data. We expect our hybrid AMN models to enhance constraint-based modeling and to prompt new biotechnological applications.